But I didn’t mention the ones I use almost every day.

And they are not specifically related to investing or asset management.

They are more general in scope.

So let’s look at 2 of my favorite Claude skills. (To install these skills, just copy-paste the URL inside Claude Cowork and ask it to install it)

Skill number 1: Grill-me

I build tools and systems to enhance my investing process. So this first skill is the grill-me skill from Matt Pocock.

It’s simple and genius. When you build or do something, be it in Cowork or Claude Code, grill-me will, well, grill you. It will start asking deep questions to fully understand exactly what you want.

Here’s an example:

In the past, I mentioned the AskUserQuestion tool, which is always a good idea: let Claude ask you lots of questions so it can build the best context.

Grill me goes a step further.

Here are 2 examples of questions it asked: Where to pull the data:

And for example: A question on the components of the dashboard itself.

And it went on and on. Grilling me until it was satisfied.

In total, Claude asked me 10 questions. Which is more than he typically would without the skill.

And what did I do?

Yep, I went for a second round.

Subscribe now

Skill number 2: Hand-off

The problem: When you’re working in Claude, sometimes you’re running out of a session. You’ve consumed your entire context window.

And even if you haven’t consumed everything, performance degrades when your session gets long. The model has trouble “remembering”.

In those 2 situations, you need to be able to stop your session cleanly, and then hand it over to a next session

(Note: There are techniques to limit token usage inside a session; we wrote about it here).

The skill that solves this is the handoff skill.

How does it work?

Everything that happened in your session is saved to a file named HANDOFF.md at the project root, which contains Goal, Current Progress, What Worked, What Didn’t Work, and Next Steps. It reads any existing HANDOFF.md first so updates chain across sessions rather than overwriting context.

In addition, it will append a “suggested skills” section telling the next agent which skills to invoke. (Because without this, the agent might not know in a fresh session).

I had a project where I was building a 10k reader. I almost used the full 1M context window with Opus, so I used /handoff

What is this assisted10kreader?

You’ll see that in a couple of weeks.

May the markets be with you, always!

Kevin

• A brain (the model)

• A harness (tools, connectors (MCP), etc.)

• An internal loop so that it works towards a goal

But what makes an agent (or you) truly powerful is their skills.

Claude skills are powerful nitbits of software that can produce a daily briefing, a full equity research report, and a lot more.

They are the backbone of your investment research.

New skills are created every day, and today I’m going to show you how to find them.

Now, if you use Claude Code, you probably know the “find skills” skill.

Yep, it’s pretty meta, but since new skills are developed every day, instead of building your own new skill, you might check if a skill already exists that covers your needs.

The problem: This “find skills” skill is only available for Claude Code, not for Claude Cowork.

Until now.

I adapted that skill so that it works inside Claude Cowork.

Findcoworkskills

This new skill is called: /findcoworkskills, and you can find it here.

Instead of building a new skill, let /findcoworkskills check the internet first.

So what does it do?

Let’s say you’re in a chat, and based on your exchange with Claude, you decide to write a skill. But before doing that, you want to know what already exists.

So you type:

In this prompt, based on your discussion, the query will be shown first. And if you agree with the query, Claude will start.

But you can also be more specific, like this:

The query Claude will use is “discounted cash flow,” and it will start searching.

You’ll see Claude doing its thing, and once it’s done, this is the report:

It’s a list of skills you can go to and install.

The list will highlight:

• If a skill is cowork-ready, that means you should be able to install it easily

• If the builder is an official vendor (e.g., Anthropic)

Important: If the skill does not list “cowork-ready,” you might still be able to install it. Just give Claude the GitHub URL and ask him to turn it into a skill that can be used in Claude Cowork. There are 2 possible replies:

• Yes, and he will just adapt the skill so that you can run it

• Yes, but not in the same form → the skill might lose some capabilities

I have never had a situation where it has given me an outright no.

The vast majority of skills are built for Claude Code. But you do see more and more skills that can also be installed on Claude desktop.

Let’s test it and look at 3 skills that have been top-rated based on GitHub stars.

Subscribe now

Skill 1: Discounted Cash Flow

Let’s start with a Cowork-ready skill.

The top-rated skill that you can install in Claude Cowork when you search for “discounted cash flow” is the DCF skill from Anthropic.

It’s included in the Claude for Financial Services package.

And you probably already installed it.

Go to customize → plugins → add marketplace → browse Anthropic sources

And select the financial services one:

Once installed, you get the DCF skill and a whole slew of other skills that sit inside that package.

And when you run it on a company, let’s say ASML, you get this:

I chose:

• Management’s growth prospects

• 10 years forecast

• Peer set: Applied Materials, Tokyo Electron

• No other data

The skill will first fetch the data it needs. This depends on the connectors you use. I use EODHD for financial data because, cost-wise, it’s the best for worldwide reach, but you could also source the data from EDGAR for US companies.

You can download the Excel files below: A DCF for ASML and a peer comparison.

Asml Dcf

26.3KB ∙ XLSX file

Download

Download

Asml Comps

13KB ∙ XLSX file

Download

Download

Skill 2: Technical analysis

I’m personally not a specialist in technical analysis. So this could be useful.

When you search for them, the skill shows you this:

The third one in this list looks interesting because it does actual computation with Python.

This skill is one part of a package of trading skills created by Stashk.

Let’s install it and see what it does.

Now you can’t just upload this skill like you normally would.

Just paste the GitHub URL into Cowork and ask it to install it there. Cowork will show you the steps you need to take.

It has a skill.md file and 2 Python scripts. So we need to upload the entire package inside Cowork.

Cowork does its analysis and decides that it can create a .skill bundle out of this:

And after a couple of minutes, Opus 4.8 produces this:

You save the skill and then run it.

Note: This skill was not cowork-ready. Apparently, it was flagged as cowork-ready because the word Claude Desktop appeared somewhere. But as I have shown you, most skills can be repackaged for cowork.

So here’s what it produces for technical analysis.

So no charts or anything. It retrieves data from yfinance, calculates key technical metrics, and then provides a conclusion.

For me, that’s enough. If you want more, browse some of the other skills in the list.

Conclusion

There are thousands of Claude skills out there. Not all are of high quality. But you now have a tool to search for some of them specifically for Claude Cowork.

I’ll use this tool to provide you with a TOP 5 or TOP 10 list of skills I consider when creating high-quality output.

In the meantime.

May the markets be with you, always!

Kevin

And in that presentation, I showed this slide:

And I just noticed that I never talked about where to source data or how to get good data for your investing research.

We’ve talked about the models and the skills; it’s time to talk about the data.

One thing to keep in mind: You can use Claude Fable 5 (well, not yet) alongside a high-quality Claude skill. It will produce rubbish if it uses bad or incomplete data.

Let me show you what my tech stack looks like. I’ll share with you my experiences with the free and paid tools out there.

Short intro on data quality

Markdown

A couple of weeks ago I talked about the second brain and how you can structure it. Again, that brain is only as good as the data you give it.

And as we discussed, LLMs do not like PDFS. (PDFs have images, have a lot of layout, and also have a lot of text that you don’t really need) They prefer Markdown files, which are just structured text files. So my own second brain at the top has 2 folders:

• RAW: These are just all the files, untouched as they come in. PDFS, DOCX, XLSX, everything.

• COMPILED: This is the result of converting all the raw data into my second brain, with structure, with relationships, ready to be read by an LLM.

Here is an example of the impact when converting a PDF annual report into a Markdown version of that report:

• PDF: 4.5 MB or 327,000 tokens

• MD: 0.3 MB or 85,000 tokens

One important caveat in this example: Text and tables are converted into the Markdown format. Images are not converted. For an annual report, I don’t really care about that. But if you have a company presentation in PDF, with lots of images and charts, you need a vision model. Luckily, Claude has a vision model. In that case, you just feed it the raw PDF.

So let’s test this and create a Claude project where in one instance, we upload the 5 PDF files, and in the other, we upload the 5 Markdown files:

Wait, the markdown is bigger?

Here’s what happens:

For a PDF, Claude will extract all text from the PDF and add it to the project. It does not extract tables (if pasted as an image). It does not extract images.

The Markdown file we added contains the same text and also includes the tables extracted as data. That’s why they consume a bit more space

But for equity research, you want the tables, the income statement, etc from the annual report. So be careful when using PDFS. Always test them, like ask Claude to extract the full income statement to see if it checks out.

Important to note: while a Claude project does not convert the images from a PDF, when you upload a PDF to Claude Chat or let Cowork read it from a folder, Claude will use its vision model to read everything! However, if your PDF is more than 100 pages long, Claude will revert to best-effort mode, meaning performance will drop.

Data access

Next to data quality is the way you access the data. And there are 2 ways:

The MCP Server

MCP stands for Model Context Protocol. You can consider it an open standard so that you can “talk” to data with any LLM. If a data platform has an MCP server, you can add the MCP inside Claude through the customize and then the connectors tab:

This allows you to use that data inside any Claude session. Bear in mind that some connectors only work on Claude Desktop. You’ll see that in your Connectors tab.

An API

API stands for Application Programming Interface. It’s another way for you to get access to data or software. The difference here is that you can access it outside of a Claude session. For example, my stock tracker tool uses an API to retrieve data outside of a Claude session.

What data?

Now let’s finish this first chapter with what kind of data you really need.

When we want to do equity research, we can just pull in all the annual reports, proxy statements, quarterly reports, and even earnings call transcripts.

Bottom line: We need textual data (to feed to the LLM) and numbers (which I like to feed to Python).

So I use different providers for different data. Here’s my setup:

Subscribe now

Tech Stack: Financial Data

Let’s first look at what the big 5 are doing. Are the big boys like Bloomberg or Factset providing connections to Claude?

Here’s the current state of affairs:

So the answer is: yes, besides Bloomberg, all the other data providers have developed some sort of access point, be it an MCP server or an API connection to Claude.

Now, for a small team like ours, these data providers would be overkill.

So here’s what my tech stack looks like:

EODHD (paid)

For worldwide financial data, I use EODHD. EODHD provides an MCP Server and an API. I use it for all my personal projects. It covers fundamental data for most countries, provides historical pricing data, and even offers a news channel.

The advantages for my use cases:

• Still affordable

• Agentic, built for agentic use

• Truly global, with about 60 exchanges

• Fundamental data going back up to 30 years

Check out EODHD here:

GO TO EODHD

EODHD provides me with the financials. Income statement, cash flows statement, balance sheet and insider ownership.

There is only one country that EODHD does not cover: Japan.

But I have another solution for that.

Edinet DB

This will give you access to Japanese financial data. And the setup is easy

Here’s how to set it up:

• Go to customize -> connectors

• Add custom connector and copy/paste: https://edinetdb.jp/mcp

Then click connect and you’ll see this window:

Now choose your google or microsoft account and you’ll be logged in.

There are many financial data providers out there. I’ve listed them below so you can check what kind of data they provide and if they already have an MCP server or API you can use inside Claude. Use the data coverage information to check if it serves your needs.

And as you can see, only 3 of them do not have an MCP Server:

• Tiingo

• Twelve data

• Marketstack

Even FRED has an MCP server if you’re interested in pulling macro data from the United States.

Financial data is great. But we also need the actual reports. The text. So that an LLM can parse it. We need all the data.

Tech Stack: Financial Filings

EDGAR

If you’re looking for filings for the United States, that’s easy. Just use EDGAR.

And it’s free.

You don’t even need to connect to an MCP server. Claude can make a request over the web, and pull any 10k. It even works inside Claude Chat:

Now, if we could install it as a connector, that would be even better; then Claude would just know it can use it.

And of course, someone built one. It’s called edgartools.

Here’s how to set it up:

• Claude.ai (web): Use the typical connector steps and copy/paste https://app.edgar.tools/mcp, then log in with Google.

• Claude Cowork: It’s a bit more complicated. Follow the steps here: https://www.edgar.tools/mcp#setup

Financial Filings.com

Financial filings are a new player on the block. They provide company filings from around the world in PDF and Markdown formats. (They do the work for us.

They have a paid API and a free MCP Server for Claude desktop.

I have been testing the MCP Server on my Microcap portfolio, and it seems to have all the data for all those companies.

That means in addition to the numerical data, you now have the narrative data.

So if you use Claude Cowork and want data for free, you can use their MCP server in combination with the finance MCP Server, and you now have quality data at your fingertips.

This can get you far.

When it comes to the pure filings, here’s what’s possible:

You can of course try the API endpoint for each country. (if a country exchange even has one).

But I still use paid products because the quality of the data in EODHD is better than the free options, and, of course, it can go up to 30 years into the past.

Tech Stack: Data and filings

And then finally, there are data providers that provide both, financials and filings in a single package. That means you would only need 1 connector:

Based on the table above, your best options are:

• Financial Modeling Prep

• Fiscal.ai

• Roic.ai

For the moment, my data needs are met.

Which data provider are you using and why?

Leave a comment

Conclusion

All quality work with Claude or any LLM starts with great data.

As you’ve seen in this article, there are a lot of free options. Paid options will provide you higher quality and more performance.

My recommendation: Do not skim on this.

I said it in the intro, but I will repeat it through a formula:

Bad Data X Great Model X Great Claude Skill = Bad Outcome

May the markets be with you, always!

Kevin

My son loves playing video games. His favorite one is Minecraft.

The goal is to kill the ender dragon. You don’t beat the dragon in one move. The fight sits at the very end of the game, and getting there is a long slog of mining, trying, and dying, again and again.

Almost everyone’s first dragon kills them, because nobody tells you it heals itself off the crystals sitting on top of those pillars. So you respawn. You bring more arrows this time. You shoot the crystals down first, then go after the dragon. Try, watch what happens, change the plan, go again.

This is an agentic loop:

1. The goal is clearly defined

2. You play and try different options to reach that goal.

3. You observe the result. As long as the goal is not reached you keep playing.

That first sentence is not a detail. You have a measurable and clear success condition.

So what’s the problem?

2. A real agent that earns its keep

The cleanest example of an agent that genuinely works is one a lot of us already have: Claude Code/Cowork.

You give it a clear and measurable goal. Build me a landing page with these three sections, a working email signup, a dark-mode toggle.

Then it goes. It writes the code, runs it, hits an error, fixes the error, runs it again, checks the toggle, and keeps looping until the thing works. You didn’t stand over each step. You set the goal and it closed the gap.

The secret of the agent is the combo of goal and loop.

Why does this work so well?

Because software comes with a dead-dragon moment built in. The tests pass or they don’t. The button works or it’s broken. Success is a hard fact you can check, so the agent always knows whether it’s finished or needs one more loop. Tests going green is its version of the dragon falling out of the sky.

That’s the secret. Agents shine where the finish line is concrete and a machine can check it without a human squinting at the screen.

When we look at what an agent is, in Claude, it can be shown like this:

Claude Code, OpenClaw, Perplexity Computer are all agent harnesses.

You use them to build and do stuff.

But building a company where each C-level suite is an agent is hype. Automating my entire workflow with agents is hype.

As I’ll show you below, you don’t need agents to fix everything.

Let’s dive deeper.

3. The skill that beats the agent

Here’s a system I run.

Every Monday at 7am a job fires on my Mac Mini. It pulls every new scrap of information on every stock I own: filings, insider trades, earnings, the news flow. It updates a file for each company, lines that file up against the thesis I wrote the day I bought the stock, and tells me one of three things.

• The thesis is intact.

• It’s under stress.

• Or it’s broken.

No babysitting. No loop, I have to sit and watch. It runs, it writes its report, I read it with my coffee.

This is a skill, not an agent. It runs the same recipe every week, and that sameness is exactly why I trust it. (you can learn how to build a skill here)

It’s predictable. When it does something odd, I can open it up and see why. It does the boring, heavy data work and hands me back the one piece that actually needs a human: what do I do about a stressed thesis?

That call is mine.

It’s my money.

I don’t need an agent here. That skill is doing all the work.

Anthropic defines agents as follows:

Let’s break down the above schematic with an example: Producing an equity research report

• Single task: You prompt to create a report on a company. The model gets called and delivers you an output. This is what most people do.

• Workflow: The goal is: A structured output of an equity research report. The model will act on a list of instructions. At a certain point, you will get your report. This is a Claude skill like the example you’ve seen above.

The agent has a loop where it observes and decides. The model will choose the actions it needs to achieve the goal. Just like my son, he will try, fail and try again.

The big difference is that it will monitor its environment and reloop if necessary.

• A build agent will produce an output just like a skill

• An optimization agent will monitor some variables and keep looping

A typical example of an optimization agent: A landing page agent which reads visitor and conversion data, and which will continue optimizing and changing the landing page until a certain conversation rate is met.

For equity research, an optimization agent is a lot more difficult to design.

4. So when do you actually need one?

Strip the hype off and the rule is simple. You need an agent when two things are true at the same time.

• The goal is concrete. You can say what “done” looks like before you start.

• And you can check “done” without yourself in the room. The system has its own dead-dragon moment.

For a while I tried to talk myself into upgrading my skills into an agent. Let it go investigate the stressed names on its own. Let it draft a sell call, set a price target. But the skill already pulls the data. It already does the evaluation. So what does the loop actually add?

More moving parts. More ways to break. A judgment call (”should I sell?”) that looks autonomous, when I’d override it half the time anyway. There’s no dead-dragon moment in selling a stock. There’s just me, deciding.

If you only learn how to build high quality skills, and maybe automate them so that they run automatically, you’ll cover probably 80-90% of your needs.

And these are not agents. These are automated structured workflows.

A true agent is much harder to build. So you would be spending a lot of time to cover 10 or 20% of your remaining needs.

5. The conclusion

Most people building agents right now are solving the wrong problem. They’re trying to automate judgment, the one thing that’s actually theirs to keep, instead of building better tools to support it.

Always make sure you steer the machine. Claude prepares. You judge. Especially in equity analysis, this is important.

So focus on building high quality skills. Add structure. Add depth. Don’t focus on building your Warren Buffett agent that runs an entire analysis without you.

It’s a waste of time.

Here’s our previous guide on how to build skills.

There are skills, and then there are institutional grade skills that you can actually use for equity analysis.

I’ll write about the difference between them in the future.

May the markets be with you, always!

Kevin

I was humbled and surprised at how many people were interested in the topic.

The venue was sold out in less than 2 weeks. 115 seats.

We could have gone for a larger space, which would allow for 210 seats. But I wanted to be close to the audience and have more interactivity.

If you were there, thanks a lot for attending!

More events will be organized soon, live and online. (or reach out to me if you’re interested)

My favorite part of such a live event is the drinks in the bar afterward. That’s where the best questions surface.

1. How can you trust anything Claude produces?

2. Does Claude really let you generate alpha in the market?

3. Or more practically: Claude burns a lot of tokens. Do you have any hacks to minimize this?

I will discuss the first 2 in separate articles because they are among the most important problems to tackle, and we have addressed them internally at Atomical Capital.

So today, we’re going to take a look at how to reduce token usage. This is especially useful if you are on the Claude Pro (20/month) plan and still want to use the Opus 4.8 model from time to time.

We’ll look at 5 things you can do and one specific solution that dates back 10,000 years.

But first things first, what do you get with a Claude Pro subscription?

What’s inside Claude Pro?

When you go from a free to a Pro subscription, you get a lot more. I highlighted the most important features in red.

More usage, of course, but 3 features I use almost daily:

1. Claude Cowork

2. Claude Research

3. Claude for Excel

Twenty dollars a month buys you a metered slice of compute, not a fixed bag of tokens. So there’s no counter in the back. It’s a bit more complicated than giving you 10M tokens and trying not to spend them all.

Pro runs on two limits: a 5-hour rolling window that opens with your first message and a weekly limit that applies across all models.

After reading the Anthropic documentation, how much you can do “varies based on message length, the length of files you attach, the length of your current conversation, and the model or feature you use.” That’s the entire official spec. It depends.

Anthropic claims roughly 45 messages per 5-hour window, and the people who measured it are clear that it only holds for short, light chat. Attach files, run long context, switch to Opus, or fire up tools, and that number drops fast.

One token pool feeds everything. Claude.ai chat, Claude Code, and Cowork all draw from the same allowance. Burn it in one, you lose it in the others.

One thing to mention, Cowork drains your usage much faster than plain Claude Chat. It runs the same agentic engine as Claude Code, and agentic work is expensive in one specific way: every turn re-sends the entire conversation history, every file read, and every tool description back into context.

In Cowork, your enemy is accumulated context, and the wrong model multiplies it. Both are fixable.

1. Use Chat for thinking, Cowork only for deliverables

This is the most important decision, and it costs nothing.

Third-party testing puts Cowork at 10 to 20 times the consumption of plain chat for comparable work. Every Cowork turn carries the full agentic overhead. So spending it on “explain this concept” or “what do you think of this thesis” is like taking a forklift to buy groceries.

I had to create the image 😉

Do your questions, brainstorms, and back-and-forth reasoning in Chat. Reserve Cowork for when you need an actual artifact: a model, a report, a formatted deck. Think inside the cheap room. Build in the expensive one.

Subscribe now

2. Lock in the right model before you start

Opus consumes roughly 3 to 5 times as many tokens as Sonnet for the same work. Anthropic confirms the direction, but doesn’t publish the exact figure.

Now remember, you cannot switch models in the middle of a single skill run. When you launch a skill in Cowork to build a report, the model is set at the start, and the entire run executes on it. The mechanical work, the Excel formatting, the reasoning, all of it runs on whatever you picked.

So pick deliberately. Default to Sonnet 4.6. It does institutional-quality research. Reserve Opus for runs where the reasoning genuinely demands it, and accept that an Opus run is an all-Opus run.

Want Opus on the hard parts only? Then break the pipeline into stages rather than a single autonomous flow. Run data-gathering and modeling on Sonnet, then start the valuation and thesis step as a separate Opus invocation. More clicks, far less burn. That’s not mid-run switching. It’s architecting the work so that cheap stages never touch Opus.

This is one of the reasons why the basic equity research skill in Claude is divided into 5 steps. You can select the appropriate model. And check the outcome before each new task.

While you’re at it: turn off extended thinking and web search when a task doesn’t need them. Lower the effort setting for routine work. These are Anthropic’s own recommendations.

3. Keep context lean and start fresh often

A focused chat or cowork session is better than a long-winded thread where you start dragging the context along.

When using Cowork and you’re using a second brain or a local folder, only point it to the files it needs. Don’t give it everything.

Start a fresh session per task. A message in a 50-turn conversation can cost many times the first message in a clean one, because the model reloads the entire thread every turn. When you switch tasks, switch conversations.

Put recurring files in a Project. Anthropic caches project documents, so referencing them only counts the new portions, not the whole file every time. For anyone working the same set of filings or models repeatedly, this is the biggest quiet saving available.

4. Use Subagents

Claude can spawn subagents to work in the background.

Now this is not a solution for your TOTAL spend limit. But it allows you to limit the usage inside 1 single chat or task.

An agent will spawn its own Claude session. The only thing that gets returned to your session is the output of that subagent.

Note: I always use subagents. It’s more efficient and allows Claude to do things for you in parallel.

5. Compress the output, the smallest lever

I promised you a solution from 10,000 years ago.

The caveman.

The caveman skill makes Claude answer in a compressed, telegraphic style, with around 65% fewer output tokens on benchmarked prompts.

Here is a response without caveman

Now we say: Talk like a caveman, and then ask the same question:

Claude really talks like a caveman! It’s useful and funny.

This second response was 22% lower in token output. But the longer the response is, the more you gain with caveman.

This is not the miracle solution. Caveman only shrinks output, and in an agentic Cowork session, output is a small slice of the total. Session-wide, you’re looking at single-digit to low-double-digit percent savings. The honest reason to use it is speed and readability, with the quota saving as a bonus. As its author puts it, cavemen don’t make the brain smaller, just the mouth.

When you install the plugin, you get multiple skills:

One nice touch is that if you forgot how exactly caveman works, just type caveman-help, and it will print a reference card that you can use:

You can grab the skill for free here. Credits go to the author of the skill.

Install it through CUSTOMIZE -> ADD PLUGIN -> MARKETPLACE -> COPY PASTE GITHUB URL

Conclusion

Pro is enough for a lot more than people assume, if you’re disciplined and you live on Sonnet. Most “I need to upgrade” frustration is a workflow problem wearing an upgrade costume: thinking in Cowork, defaulting to Opus, hauling a 50-message thread behind every request.

But I’m not going to lie. There’s a real ceiling.

I use Claude daily and run multiple equity research skills. That means I use Opus 4.8 often, and when you do, the pro plan won’t cut it. That’s why I personally have a max plan.

But this is my job, my passion. And the ROI I get on that max plan is high

So the real decision tree is short.

• Light and occasional, stay on Pro and run Sonnet.

• Heavy and Opus-dependent, Max 5x.

• All-day agentic work, Max 20x.

And before you pay anyone, run a representative report and watch both bars in Settings > Usage.

That’s your actual number. Nobody publishes it, and it beats every estimate in this article, including mine.

So, any other ideas to avoid unnecessary token spend?

Leave a comment

May the markets be with you.

Kevin

We talked about how to build them.

And how you should create your custom skills to enhance your unique investing process.

But Claude has released skills of their own, for free!

Anthropic engineers worked with Accenture and banks like JPMorgan to develop skills and full agentic workflows.

The only problem is that they are only natively available for the enterprise and teams plans.

But there’s a workaround for installing them. I’ll show you below.

A recap on skills

A skill is a recipe. It has a set of instructions. It can compact multiple prompts together. You can add Python scripts.

Essentially, it’s a small piece of software. It has inputs, and it will deliver an output.

Here’s an example of the output of a company lifecycle skill I use. It fetches data, uses the lifecycle framework, and then outputs a report that:

• Tells us where the company sits in its lifecycle

• Proposes the appropriate valuation methods you should use

• Verifies whether, as Professor Damodaran says, the company is acting its age (from a capital allocation perspective).

The skill creates a 2-page report.

A visual representation of where the company sits in its lifecycle:

And a 2-part analysis with valuation methods to use or avoid:

You can imagine building 10 to 20 skills like these, and then chaining them together to get a full equity research report.

That’s what Anthropic did.

The only caveat: it’s tailored to a sell-side research firm.

But before we go deep, let me show you how to apply these new skills.

Install the skill library

1. Open the Claude desktop app

2. Go to Cowork

3. Click on Customize

Now this window will appear:

Pick Create new skills.

Now, logically, you would add a new skill, but that’s not the correct path here.

You need to click on the + sign to add a new plugin.

Then you select “ Create a plugin and add marketplace.

I know, it’s a bit weird.

Now this window will appear

These skills do not appear in the official marketplace. We need to add them from a GitHub repository.

Now you paste this URL in: https://github.com/anthropics/financial-services-plugins

And click the sync button.

You now have access to a set of new skills:

Now let’s take a look at one of them and see if they are actually useful.

Subscribe now

Claude Skill: initiating-coverage

When a sell-side research firm starts covering a company, it will publish a detailed, well-written report on it.

The company will, of course, have a price target.

I tested this with Claude Cowork on a company called Hyloris Pharmaceuticals.

Just for context, Hyloris develops medicines in the drug repurposing market. They have some compounds already commercialized and others still in the pipeline. The charts below will make more sense.

Just open a task and type

Claude will ask you for additional information and will get to work. It uses its questionnaire tool to make it easy for you:

The workflow will be split into 5 discrete tasks. It took Claude Cowork with Opus 4.8 about 45 minutes to create the full report. It took me about 10 minutes just to check in from time to time and go through the different tasks.

The result: A 45-page, highly detailed report:

Riddled with financial data and charts.

A couple of my favorite charts it generated:

A timeline with milestones

A competitive position matrix

A projected revenue chart, by compound

Football Field valuation

The valuation part is interesting. The terminal value is about 85% of the total DCF valuation. So a slight change in which perpetuity rate you choose will have a big impact on the equity value.

Now, if you remember, I ran my own lifecycle skill at the start of this article.

And for Hyloris, it says:

So, in this case, I would simply ignore the DCF output. It’s just too sensitive for this early-stage company.

However, the report did do a peer comparison valuation and even correct the EV/Sales multiple because Hyloris has a higher gross margin than its peers:

Now I think the idea of using EV/sales is ok. EV/Gross profits would be even better because it automatically includes gross margins.

So yeah, room for improvement. Always analyze the valuation outcome in detail, no matter which skill you use.

Positive: It explains how it arrived at the number, so it’s easy to verify.

Practical considerations:

A task in Cowork is limited to 1M tokens. I had to use 2 tasks to complete the full report.

You can always use the /context skill to see your current usage and where the tokens went. And as a reminder, when you just start a task, it will already consume 5% because of all the prompts and tools it preloads:

Since this skill is divided into 5 parts, the first 3 were in task 1 and the remaining 2 in task 2.

Since we’re working in Cowork, always assign a folder where it can work. This makes it easy to switch from the first to the second task without losing context.

I ran this on Opus 4.8 High. I ran into a session limit right when the report was finished, but I had been using Opus for other tasks before. So I do think that, when using the best model, you should be able to get a full report on 2, maybe 3 companies.

For context: I run the max 5x plan. If you use the 20x plan, you can, of course, get more out of it.

Assessment of the quality of the report

The overall quality of the report is very high.

Company overview and history. Go to market and product lineup. Competitive positioning and assessment. Market potential and a full scenario analysis.

This is to be expected. The data is retrieved from annual reports, and I did a fact check on the financial data; it was spot on. No errors detected.

Now, when it comes to valuation and scenario analysis, I think it did a good job, but I personally do not like price targets.

It has identified a catalyst that would prompt them to reassess their view, but for now, they maintain a sell rating with a price target of 3.8, versus the current market price of 5.

This is an initiation report, not a thesis. This is not a “why I buy this stock document”, and these are my kill conditions and selling rules.

A thesis is needed if you want to create an automated tracking system, as we have shown here. However, this research report could serve as the basis for a thesis.

How to use this skill

The best skill is one that aligns with your own process, how you look at companies, and how you research them.

I think this skill is a great starting point to make it your own.

A couple of things I would change for myself are:

• Losing the price targets

• Adding a more in-depth business model analysis

• Add my lifecycle skill so that the research is adapted to where the company sits in its lifecycle.

We only looked at one skill. Notice that if you installed the full plugin, you now have 15 new skills to work with.

In the future, I’ll test each one of them with Opus and Sonnet to see if there is a big difference.

So now you have the skills at your disposal the pros use.

Play with them. Test them. Push them to their limits.

Thanks for reading.

May the markets be with you!

Kevin

It’s still empty.

He’s 19. He’s in his second year of a bachelor’s degree in economics. He knows what a DCF is and the difference between a limit order and a market order, but he’s never bought a single share.

Not one.

That worried me. If I could go back in time, I would invest and compound every dollar I had.

So we made a deal.

Every week, we sit down for one hour. No more. The question we’re testing isn’t whether he can learn to invest; he can, anyone can.

The question is whether AI can take him from an empty account to his first purchase within four weeks and build enough conviction along the way for it to stick.

Part 1: Why, why, why?

We could have started with some interesting stock ideas. We could have started with Claude on how it can help him.

But I chose to ask one simple question: Why do you want to invest?

I first shared my own reasons all those years ago: I made a decent living. My wife hated her job and wanted to take care of the family. So she quit.

But her quitting has a significant impact on our pension. So I ran the numbers to see how much capital I would need to build to compensate for the pension gap.

I worked out the number I had to compound at to close that gap. It’s not a small number. Whatever yours is, it determines your entire investing strategy.

Every person has their own number.

So I asked him again: Why do you want to invest? And what’s your number?

He paused. He didn’t really have one.

At 19, that’s normal. But I waited. Gave him time to think.

He would like to own a house later on. And if he could already build some capital on the day he graduates, that would be a benefit. He would have options.

So now he had his first time horizon: 4 years.

But that’s not long enough when you invest in the stock market.

He agreed. The goal was to invest for a much longer period. Graduation could be a first milestone.

Part 2: Start with what you know

Based on his approach, we decided to build a portfolio that would be a mix of ETFs and individual stocks. Exact composition unknown at this time.

My goal here is simple: To get him excited about investing in the stock market, not get too technical.

Now I wanted this investing thing to be grounded in his real-world experience. When investing becomes fun and real, people tend to stick with it. When it’s abstract, not so much.

So I borrowed Peter Lynch’s approach and asked:

Which brands do you like? What do you see around you? What are you a fan of?

He responded instantly: Uniqlo

Uni-what?

Uniqlo is a clothing brand I had never heard of. A quick Google search revealed that it’s owned by the company Fast Retailing. (Ticker: 9983)

I pulled up fiscal.ai and showed him the basic stats like revenue, earnings, and free cash flow per share:

This thing is on a roll.

He said it’s durable clothing at a reasonable price.

Nice. So I asked him how this business makes money?

Besides that, they make money on the clothes they sell; he couldn’t really answer, not in detail. No worries.

Time for AI to shine: The Feynman prompt

The Feynman prompt explains any business in simple terms so that anyone can understand how it operates. You can find more details on it here.

I gave him the prompt, but you can also turn it into a skill. I use it like this:

And Claude started working on it:

Here’s what we learned:

Fast Retailing is a Japanese apparel company that designs, manufactures, and sells simple, high-quality everyday clothes (mostly under the UNIQLO brand) to anyone, anywhere in the world.

Fast Retailing runs an apparel model called SPA: Specialty store retailer of Private-label Apparel. In plain English, that means they own every step of the clothing pipeline themselves.

Here is how the loop works:

1. Design in-house. Their R&D centers in Tokyo, New York, Paris, Shanghai, and Los Angeles design every product. No outside brands, no licenses.

2. Buy fabric directly. They negotiate large-volume contracts directly with material suppliers like Japan’s Toray Industries, the partner that co-developed Heattech.

3. Manufacture at partner factories. Production is outsourced to long-term partner factories in Vietnam, Bangladesh, China, Indonesia, and India. Many products are made in lots of around 1 million units per item.

4. Sell only through their own stores and websites. No wholesale partners, no department-store concessions. Every yen of margin stays in the building.

Here’s what they choose not to do.

Most apparel retailers chase fast-moving fashion trends and get punished when last season’s product goes unsold. Fast Retailing deliberately sells timeless basics

Uniqlo is the opposite of Zara.

Zara is fast fashion. New designs hit stores in two weeks. The whole point is novelty.

Uniqlo plans collections six to twelve months in advance. Same basics, year after year, in better and better materials. They have a proprietary partnership with Toray, a Japanese chemical company, to produce fabrics that nobody else can buy.

That’s the moat: Not speed, but durability, quality, and patience.

Subscribe now

Part 3: Great business, great investment?

I saw he was getting excited now.

We found a brand he adored, and based on the numbers and business model, it’s a phenomenal company.

Yes.

“I’ve found the first stock I can buy”.

But then I asked him a question: It’s a great company, but is it listed at a great price?

He had no idea.

So we set out to reopen fiscal.ai and review multiples and growth rates.

And since he already learned about a discounted cash flow, I told him multiples are a shortcut to a full DCF. As long as you respect that, you are allowed to use multiples as a rule of thumb.

To make it simple, we looked at the historical PEG ratio.

A PEG of 3 is expensive.

There was an opportunity in September 2023 (PEG of 0.7) to buy this stock at a pretty low price. You would have more than doubled your money in 3 years.

The market has priced this stock in.

My advice to him: Open a brokerage account and add it to a watchlist. Set a price alert. Move to another stock.

This is a stellar business, but a great business can be a bad investment.

Conclusion

So that was it for week 1. My nephew now has access to fiscal.ai and Claude with some custom prompts to look at different businesses.

He learned:

• That investing serves a purpose, a why

• That he can understand any business by using Claude

• That a great business does not always equal a great investment

In the coming weeks, we’ll go deeper into the business, look at some ETFs, and see how this goes.

My nephew is in the driving seat. I’m just there to guide him.

If you started investing as a teenager or in your twenties, what’s the one thing you wish someone had shown you in your first month? (My nephew reads all the comments.)

Leave a comment

Keep learning AI, keep building!

Kevin

Over the past 6 months, I’ve built tools to hunt for new ideas, track existing holdings, and counterpunch each thesis.

We talked about using Claude projects. Building new skills. And a first step into automation by using Claude Cowork.

But that’s not what creates real leverage when you’re investing in stocks.

These are useful tools, running on top of powerful models.

But the models will get commoditized.

And in a couple of years, everybody has these tools at their disposal.

But each investor has his edge. What makes him unique.

And that is your second brain. A living brain. That’s what you need.

Maybe your current situation looks like this:

Three months ago, you spent ninety minutes walking Claude through a thesis on a Japanese regional bank. The conversation was excellent. Claude pushed back on your assumptions, surfaced a regulatory angle you’d missed, helped you size the position.

Today you open a new chat to update the thesis. Claude knows your name. It knows your profession. It does not know the bank, the thesis, or a single thing you concluded that afternoon.

You start typing the backstory.

Again.

This is the central problem with AI as it exists today.

The frontier models are extraordinary.

But they forget a lot.

Every chat begins from roughly the same blank page, and every insight you produce evaporates the moment you close the tab.

They need context.

I’ll show you how to solve this, from simple to advanced. Here’s the outline for today:

1. What Claude actually remembers

2. Where this falls apart for serious work

3. Taking a mental shift before going further

4. The solution from simple to more advanced:

   1. Claude projects

   2. Claude connectors

   3. Claude code

5. Beware of these mistakes

Subscribe now

1. What Claude actually remembers.

It’s not that Claude doesn’t remember anything.

In fact, Claude has 3 memory mechanisms.

1.1. The memory synthesis

Go to settings and capabilities.

You’ll see this:

Claude reads your chats and writes a third-person profile for you. What you work on, what frameworks you use. This profile refreshes every 24 hours and will get pre-loaded into every new chat.

Click on View and manage memory to see what it has written about you.

You can adapt it in 2 ways:

• Adapt the text itself (after clicking on view)

• In any chat, tell Claude: “Remember that I… focus on small caps with insider ownership” -> It will add this to the memory file

1.2. Search past chats

This is actually useful.

You can ask Claude in any chat: “What did we conclude on that regional Japanese bank?”

Here’s a an example of a company I looked at recently:

And in the first image we showed you, you can toggle this off and on. It’s on by default.

Note: This is for paid plans only.

1.3. Claude Projects

We talked about it in this article.

You can build a dedicated memory pool for a project. Like this:

But everything you put in here is walled off from everything else. An analysis on company A will not bleed over into company B.

This can be useful, but maybe you want some part of that information to bleed over. That is not possible with projects.

If you do nothing else after reading this section, set up three or four Projects today and spend ten minutes editing your memory. That alone puts you ahead of 90% of Claude users.

2. Where this falls apart for serious work

Claude’s native memory is fine to have some sort of a pre-recorded introduction to each chat.

It’s not infrastructure that evolves for serious thinking.

Here are 3 reasons why it might hurt us as investors:

• It’s a summary. Not an archive: Your four-hour deep dive on a small Japanese bank is boiled down into a condensed format. Yes you can search the chat, but you need to know what to search for.

• There is no structural retrieval: “Show me every company I’ve researched over the past 5 years with insider ownership above 50%”. You won’t get an answer. There’s not structure in the data

• There is no timeline to your thinking: When did you change your time on a position? What did you think in February and no longer holds true today, and why? The memory synthesis is just updated every day, but you lose the journey.

But there are ways to solve this, but before we get to that, we need to change our process.

Subscribe now

3. Taking a mental shift before going further

Writing down notes isn’t new.

• Niklas Luhmann’s Zettelkasten

• Tiago Forte’s Building a Second Brain and his CODE method (Capture, Organize, Distill, Express)

• Andy Matuschak’s evergreen notes

There’s a 50-year tradition of structured personal knowledge that the AI era didn’t invent.

What changed is the reader.

For most of the past decades, the reader was a future you, opening the notebook to find what past-you wrote.

The bar for keeping a note was “will I re-read this?” and let’s be honest, most people didn’t bother to re-read.

Now the reader is an LLM that will happily ingest 100% of your archive in seconds.

Keeping notes has become much more interesting. But there’s a caveat. The bar for structuring notes so an LLM can find and use it has risen.

The time has come to build an evolving second brain. Once you do, any output the LLM will give you will be of higher quality.

The format?

Markdown files.

Plain text, readable in any editor for the next 40 years. It’s an open-format, so no vendor can take it away. Markdown files in folders will be readable by whatever model dominates 2030.

4. The solution, from simple to more advanced

Building a second brain can be done in many ways.

Pick one of these three based on how much time you want to spend on it, and what kind of result you desire.

My recommendation: Don’t overthink it.

It’s better to start small and simple than trying to concoct the most advanced second brain known to man. If you’re not doing any of these today, and decide on just one, you’ll get better results with Claude tomorrow.

4.1 Claude Projects (no setup)

I’m not going to rehash the details of what we wrote here. And although this is not a full second brain (each project is part of the full brain and cannot talk to one another), you can imagine creating a project called: Equity research. And then upload your frameworks and annual reports or data inside it.

You could probably analyze a couple of companies this way, in one single project.

If you’ve never done this before, this is the fastest way to get better results.

4.2 Claude Connectors (the no-terminal upgrade)

This is where Claude stops being a chatbot and starts acting like an analyst with access to your filing cabinet.

Start building your second brain inside your Google Drive

Keep a folder of markdown notes. You could use the following structure:

• /Secondbrain for the main structure

• /research with per company notes

• /wiki for evergreen concepts and investing frameworks

• /writings for draft or other pieces

• /journals for daily or weekly logs

• /inbox with unfiltered information

You can decide this structure yourself. You could for example have one folder per company, and inside each folder, have a structure like:

• Notes

• Earnings Reports

• Thesis

It’s up to you.

There is no wrong or right answer. The only thing you need is to decide on a structure and stay consistent with your data capture in the future.

But do make sure to take care of how you name the files. Here’s what the LLM loves:

Never use spaces. Start with a data. Describe what it is.

Be consistent.

Note: If your current archive holds pdfs, word files, and what not, you can just upload them to Claude and ask him to convert them into md files. Or you can use MarkitDown which is a tool specifically designed for this.

I do recommend using some filing system, but I prefer what the current master of AI is doing (Andrej Karpathy).

Build a logical filing structure of markdown files. Feed your LLM the right markdown files based on the task at hand.The problem with summarizing information is you’re losing context. And our goal is to maximize context.

The goal is that your second brain lives. And evolves over time. You will be adding and maybe subtracting information all the time.

Now let’s say you have a google drive with your second brain inside it.

Now you go to Claude -> Customize -> Connectors and connect your google drive:

As you can see in the picture, if your brain lives in Notion, you can connect Notion too.

But there’s one problem. Claude will read your brain through the Notion MCP Connector. The readability of the data by Claude will be less clean as opposed to pure markdown files.

So what’s one thing you can do once you’ve connected google drive:

When you do research on a company, end the Claude Chat with:

Save this conversation to /your  structure in my Drive.

Or you can also say:

Read everything in /research from the last month and tell me which thesis are under pressure.

This setup is a lot more powerful than not using anything.

But the tradeoff compared to our final solution below is real. Connectors are less powerful than the Claude Code setup. Claude web does not read directly against local files. It’s a bit slower.

But you now have a wiki synced over any device, whether you’re using Claude through your browser, on the desktop app or on your mobile.

If you implement this, you’re already gone a long way. Now let’s get to the final setup.

4.3 Claude Code (max power, terminal required)

For readers who’ll open a terminal: this is the strongest setup available right now. Claude Code is Anthropic’s command-line agent.

One install and then point claude inside your SecondBrain folder.

The second brain folder you need to build is the same as in part 2. And yes, you can use Obsidian to structure this, but you don’t absolutely need it. Claude can just look through your folders without it. Obsidian allows you to easily edit the markdown files, and visualize the relationships between the different files. Claude doesn’t need it per se.

If you launch Claude Code inside your second brain folder, it now has access to your entire brain.

Here’s how memory is now structured in a typical Claude Code setup:

• Global Claude.md file -> This is the file that is read at every Claude session. Antrophic recommends to keep it lean, like about 200 lines. Usually it will contain some do and not do’s for coding. But you can add rules related to investing if needed. This md file is the most similar to the memory file Claude.ai creates himself.

A simple example:

# Who I am

I write the “Invest with AI” newsletter. Plain prose, active voice.

Banned: delve, tapestry, navigate.

# Frameworks I use

- Munger inversion: ask what kills the thesis first

- Reverse DCF: solve for growth/margin the price implies

- Position sizing: Kelly-informed, never >5% on a single name

# Defaults

- Cite sources with URLs

- Default currency USD

- For numerical claims, show the calculation

• Project level Claude.md file -> Let’s say you have your second brain folder. Your second brain is focused on investing research. You can for example create a claude md file in this folder that looks like this:

# How I research companies in this folder

# Folder map
/wiki/companies/{TICKER}.md  — per-company accumulated research
/wiki/frameworks/            — analytical frameworks (reverse-dcf, kelly, etc.)
/wiki/concepts/              — evergreen analytical concepts
/journals/decisions.md       — log of position changes with reasoning
/inbox/                      — capture, processed Sundays
/reference/templates/        — file templates

# Company research workflow
When I research a new or existing company:
1. Check if /wiki/companies/{TICKER}.md exists. If yes, read it first.
2. If no, create it from /reference/templates/company.md.
3. Update the file as we work — don't summarize away prior thinking, append.
4. End the session by appending a dated note to the bottom: 
   "Session {date}: what we covered, what's still open, what would change my view."

# Frameworks I default to
- Munger inversion: ask what kills the thesis first
- Reverse DCF: solve for the growth/margin assumptions the price implies
- Margin of safety: minimum 30% discount to estimated intrinsic value
- Position sizing: Kelly-informed, never >5% on a single name

# Conventions
- Cite sources with URLs for all factual claims
- Show calculations for numerical claims
- Default currency USD, default fiscal calendar calendar-year
- Push back on weak arguments instead of agreeing
- When I change my view on a position, also log it to /journals/decisions.md

That means each chat you open inside this folder will get this context loaded in.

So now your chat already has some context. It has not yet touched the files inside the second brain.

You can open Claude Code within the second brain folder, or point it towards the second brain. In any case, Claude can access it and read anything from it, depending on what you’re asking in your Claude session.

Now there’s one thing that is important.

Your journal.

When you launch a project in Claude Code, it’s common practice to keep a memory.md file and a progress.md file.

This is basically a running memory of what has happened and what still needs to happen in the future.

It’s the final piece of the puzzle in your second brain setup.

In the example above, there is a decisions.md file and weekly reviews file. This makes the brain come alive.

What’s important is to design your own system.

That system will outlive Claude or any LLM, and might be more valuable than the tools you use.

5. Beware of these mistakes

Don’t over-engineer. A 17-folder PARA hierarchy with a custom frontmatter schema you abandon in three weeks is worse than five folders and a habit you actually keep. Start with the starter blueprint above. Let structure emerge from real use. Don’t overthink it. Keep it simple and iterate.

Capture without distillation is simply hoarding. Capture information, but don’t forget to distill it. Create those md files. Weekly update your memory files. A stale brain won’t work. Note: There are ways to automate this so Claude does this automatically, but that’s food for another article.

A stale CLAUDE.md is worse than no CLAUDE.md. A file describing a thesis or workflow you’ve moved on from actively misleads Claude every session. Prune monthly. Run /memory to see what Claude has learned on its own and delete the redundancies from CLAUDE.md.

Plain text is the 10-year bet. Every format choice carries vendor risk except markdown. If you start in Notion’s proprietary blocks, you’ll regret it the day you want to try a different tool. Markdown files in folders will be readable by whatever model dominates 2030.

Conclusion

I hope I was able to convince you that building your second brain is worth more than the tools you use.

Whether Claude, ChatGPT or another model wins in the future does not matter.

As long as you’ve created and maintained your second brain, you can plug in any LLM you want.

If you already built your second brain, let me know in the comments how you did it!

Leave a comment

Keep learning AI. Keep building!

Kevin (Atomical Capital)

All of them.

And that’s a problem if you use them for stock analysis.

A hallucination is when an LLM generates a confident-sounding answer that isn’t grounded in real data. It makes up the answer. For stock analysis, that’s the difference between a real revenue figure and a fabricated lie.

Here’s the Artificial Analysis hallucination benchmark data:

Grok 4.2 scores best. In 17% of the questions asked, the response should have been omitted rather than made up.

GPT-5.5 has a high score of 86%. (That’s pretty bad!)

And something that might surprise you: Claude 4.5 Haiku with reasoning hallucinates less (26%) than Opus 4.7 in adaptive mode (36%).

But how is it even feasible to use these models if a quarter of the time, they are just making stuff up?

Before diving into what you can do. GPT5.5 and Opus 4.7, from a pure intelligence standpoint, are the strongest models today.

For complex financial analysis, use a frontier model. When it comes to finance tasks, Opus still reigns supreme:

Secondly, this hallucination-rate test covers 6000 questions, in which the LLM answers based on its pretraining. That’s an important nuance!

Here are 6 things you can do to reduce the hallucination rate. We’re going to use Claude as an example, but you can apply these principles to any model you use.

TL;DR: six ways to cut LLM hallucinations for stock research:

1. Load full context (Claude Projects, CSV/markdown over PDF)

2. Hardcode “don’t fabricate” in system instructions

3. Always run reasoning mode

4. Demand sources for every claim

5. Color-code outputs by data confidence

6. Run two models in parallel (”Council mode”) and reconcile

1. Add lots of context

Make sure you give it all the context it needs (don’t let it respond based solely on its pre-training).

The bad scores above are largely due to this.

Use Claude Projects. Or just upload your data into a chat. Here are some basic rules:

• LLMs have trouble with PDFS.

• LLMs love CSVs, text files, or markdown files.

So it’s not only what data you provide, but it’s also how you provide that data.

For example, when I research a microcap, I drop the last five annual reports as markdown plus a single CSV with five years of P&L. Claude reads that combination and stops inventing numbers.

2. Tell it to be honest

Hardcode it in its instructions (for the project or general instructions) that it cannot fabricate stuff. If it doesn’t know, you prefer it tells you that.

3. Always think

Always use reasoning mode

Claude now has adaptive mode, so based on the question and context, it will switch between fast and reasoning modes. That means in your prompt, you need to ask it to reason.

The trigger phrase that activates reasoning is asking Claude to “think step by step” or “reason through this carefully.” Adaptive mode picks up that cue and switches.

Too bad Claude had to push this adaptive form upon us. I preferred selecting the thinking mode myself.

Subscribe now

4. Ask for sources

In the instructions, ask for sources

When the LLM produces a result, it should always be explicit where it got its data. Ask this explicitly in the system prompt or the single prompt you use.

If the model gives you a weird answer, at least you can fact-check it.

Reject any answer where the source is “general knowledge” or unspecified. Treat that as the model guessing.

5. Color-code your LLM outputs.

Color coding

The output needs to be presented in a certain way.

Ask the LLM to cross-check any data pulled from an annual report with the financial data you got from a database. Green means up to 3% accurate. Yellow means more than 10% drift. Red means danger: Possible error.

Add this to the prompt or to the project instructions.

It’s a simple visual hack.

6. Council mode: cross-check between 2 models

We already mentioned Council mode, in which the same problem is presented to 2 models that analyze it independently. Then you compare the similarities and the differences.

Council mode = giving the same prompt to two independent models, then comparing the two answers for agreement and contradictions. Where they agree, you trust. Where they disagree, you dig deeper.

Perplexity Pro can do this automatically, but it will cost you 100 USD/month.

Here’s an example of what the output looks like:

But if you have a 20 USD ChatGPT subscription and a 20 USD Claude subscription, you can prompt them the same way, or create a project in both, and then you’ll have 2 outputs.

Copy paste 1 output into the other chat window, and ask the LLM to spot the differences. You can even add the output above from Perplexity to it so that the LLM knows what kind of format you want it to give you.

This last method is still the best way.

I’m building a skill on this to make it easier for you, so stay tuned.

Keep learning AI. Keep building!

Kevin

Share

I have been using Claude Code heavily to develop investing tools to track my investments and research a slew of companies.

But I can understand that most of you don’t want to open a terminal, start coding, and spend a couple of hours without a guaranteed result.

Claude Cowork is sort of Claude Code, but for non-developers.

But the distinction is the following:

• Claude Code is for building and automating. It’s the most versatile of the bunch but comes with a steeper learning curve.

• Claude Cowork also lets you automate tasks, sort of

That’s what we’re diving into today.

How to use Claude Cowork to analyze stocks.

As Claude himself so neatly puts it:

• Chat is for thinking

• Code is for building

• Claude Cowork is for doing

What Claude Cowork can do that Claude projects can’t.

We talked about Claude’s projects in a previous article, but Cowork goes beyond its capabilities.

Here are some examples of what Cowork can do that Projects cannot:

• Read and write files in a folder

• Build Excel files with formulas

• Scheduled tasks that run things on auto

• Use plugins: Like financial or equity research

So what would this look like in practice? Let’s build a stock analysis and first valuation model within Cowork. We’ll take the company Adyen as an example. (for our full process, you can check it here)

Step 1: Set up the scaffolding

We’re going to give Claude Cowork access to a folder on our computer. It can create, delete, and modify files.

So we created a stock research folder and added it to the chat.

The next step is to create a structure for each company that Cowork will use. Here’s an example of a structure for Adyen that will live on your computer.

ADYEY/

├── filings/ (annual reports + press releases)

├── transcripts/ (earnings calls)

├── charts/ (weekly chart)

├── model/

│ ├── adyen_financials.csv

│ └── ADYEY_model.xlsx ← real Excel model

└── memo/

├── 01_business_overview.md

├── 02_technical_read.md

├── 03_thesis.docx ← investment thesis

└── 04_one_pager.pdf

It’s easy.

Just ask Cowork to set up a folder structure, then copy and paste the structure above.

In a matter of seconds, Cowork will build this:

Step 2: Download the data

We now have to fill out the folders. That means downloading annual reports and earnings transcripts and getting our hands on financial data.

For financials, you have 2 options:

• Free: Use finance (see prompt below)

• Paid: Connect your data provider of choice

Go to Customize, Connectors, and you can find everything finance-related by filtering

For the purposes of this article, I’m going to assume you don’t have a data connector, so we’ll use yfinance to illustrate.

Now just paste the following prompt and let Cowork do its magic:

You have access to my ADYEY research folder. Please gather the following source material for Adyen (ticker ADYEY, primary listing ADYEN.AS on Euronext Amsterdam) and save everything into the correct subfolders. Work through the steps in order and tell me if anything fails so I can help.

1. Annual reports (last 5 years) Go to https://www.adyen.com/investor-relations and download Adyen’s five most recent annual reports as PDFs. Save them in filings/ using the naming format ADYEN_Annual_Report_YYYY.pdf.

2. Quarterly / half-year reports (last 4) From the same investor relations page, download Adyen’s four most recent shareholder letters or half-year / trading update reports. Save them in filings/ as ADYEN_Report_YYYY_HX.pdf or ADYEN_Trading_Update_YYYY_HX.pdf depending on the document type. Note: Adyen reports semi-annually, not quarterly, so you’ll mostly find H1 and FY reports plus trading updates — grab the four most recent of whatever is available.

3. Earnings call transcripts (last 4) Find the four most recent Adyen earnings call transcripts. Try Adyen’s investor relations page first, then fall back to public sources like Seeking Alpha, Motley Fool, or Roic.ai. Save each one as plain text in transcripts/ using the format ADYEN_Transcript_YYYY_HX.txt. If a transcript isn’t publicly available, note which one is missing in a file called transcripts/README.md.

4. Financial data via yfinance (last 5 years) In the Python sandbox, install yfinance if needed and pull the following for ticker ADYEY:

• 5 years of daily price history (Open, High, Low, Close, Adjusted Close, Volume)

• Annual income statement

• Annual balance sheet

• Annual cash flow statement

• Quarterly income statement, balance sheet, and cash flow statement

• Key info dictionary (market cap, shares outstanding, currency, sector, etc.)

Save each dataset as a separate CSV in model/:

• ADYEY_prices_5y.csv

• ADYEY_income_annual.csv

• ADYEY_balance_annual.csv

• ADYEY_cashflow_annual.csv

• ADYEY_income_quarterly.csv

• ADYEY_balance_quarterly.csv

• ADYEY_cashflow_quarterly.csv

• ADYEY_info.csv

5. Summary When you’re finished, give me a short summary listing exactly which files you saved, which (if any) you couldn’t find, and flag any data quality issues you noticed in the yfinance pull (missing years, currency mismatches, etc.). Do not start analysis yet — this step is just data collection.

You can see Cowork doing this exactly as asked. It takes about 5 minutes to fetch all the data.

Here’s a glimpse of what’s inside the annual reports folder:

Note: It took 5 minutes, mainly because it was trying to fetch the earnings transcripts. If you want these faster and more reliably, you’ll need a paid provider, such as Financial Modeling Prep, which has an MCP server. Cowork will then be able to use that one.

So we did the setup, we have the data, let’s continue.

Step 3: Let Claude understand the business

Now that we have everything set, Claude needs to ingest the actual information.

Here’s the prompt:

“Read all PDFs in filings/. Write a one-page business overview of Adyen in memo/01_business_overview.md covering: what they do, how they make money (take rate, processed volume, net revenue), main segments and geographies, top customers and customer concentration, competitive position vs. Stripe / PayPal / Block / Worldline, and the three biggest structural risks. Use plain English, no jargon.”

It generated this business overview markdown file, which I think provides a good basic coverage of what Adyen does.

Step 4: Turn the yfinance CSV file into an Excel

The data from yfinance just sits in several CSV files. We want to build a solid Excel out of it.

Here’s the prompt:

“Using the CSVs in model/, create model/ADYEY_historicals.xlsx with three tabs: Income Statement, Balance Sheet, Cash Flow. Show the last 5 years side by side. Below each statement, add a ‘Key Ratios’ block calculating: revenue growth YoY, gross margin, EBITDA margin, FCF margin, net cash position, and ROIC. Format numbers in € millions, one decimal. Flag any line item where yfinance data looks incomplete or inconsistent.”

So Cowork created an Excel, and a summary of what it did, and you immediately see the power of Opus 4.6

So, per point 3, we ask Claude to update the Excel file and include the float.

It updated the existing Excel with this:

Despite the minor naming error, this is important if we want to work with the correct enterprise value and free cash flow later on.

Claude did warn us to cross-check the data against the annual reports, as the actual cash calculator is an approximation based on payables.

Of course, we’re not going to do that. We just asked Claude to do it for us.

So its approximation was only slightly off. It then proceeded to update the Excel.

Subscribe now

Step 5: Build a valuation model

We’ll build a simple 5-year DCF and comparables tab with the following prompt:

“Add two new tabs to model/ADYEY_historicals.xlsx: ‘DCF’ and ‘Comps’.

DCF tab: 5-year forecast with three scenarios in separate columns — Bear (10% revenue CAGR, EBITDA margin flat at current level), Base (18% CAGR, margin expanding to 55% by year 5), Bull (25% CAGR, margin to 60%). Assume 22% tax rate, 9% WACC, 3% terminal growth. Calculate enterprise value, equity value, fair value per share in EUR, fair value per ADYEY ADR in USD (use a sensible EUR/USD rate and note it), and upside vs. today’s price.

Comps tab: Pull in Stripe (note: private, use latest funding round valuation), Block (SQ), PayPal (PYPL), Nuvei (NVEI), and Worldline (WLN.PA). For each, show market cap, EV, revenue, EBITDA, EV/Revenue, EV/EBITDA, and revenue growth. Apply Adyen-appropriate multiples to get a comps-based fair value.

All formulas should be live, not hardcoded. Add a ‘Summary’ row at the top of each tab showing the fair value conclusion.”

Now, if you look at this step, you could also fetch the data from the comparables in the previous data step. That’s a choice.

So the DCF Excel includes actual formulas. So you can adjust the WACC or the terminal growth rate to better align with your own assumptions.

Based on this analysis, the overall case is bullish towards Adyen.

Let’s look at the comparables table:

Stripe is the most comparable company. Of course, Stripe is valued higher, and we need to be careful as it’s not a public company.

Step 6: Create the investment memo

Now, let’s create the final investment memo from all of the above.

Use the following prompt:

“Read memo/01_business_overview.md, memo/02_key_questions.md, and the DCF + Comps tabs in model/ADYEY_historicals.xlsx. Write a 2-page investment thesis in memo/03_thesis.docx with these sections:

1. Thesis in one paragraph

2. Why now (catalysts)

3. Key drivers of the bull case

4. Valuation summary (base case fair value, upside, comps check)

5. Key risks

6. What would make me wrong (the anti-thesis)

7. Position sizing and suggested entry range

Tone: professional but conversational, like a buy-side analyst writing to a PM. No hype, no fluff.”

Now, of course, you can use your own framework for an investment memo. Even better would be to provide your own MD file with your typical memo framework.

You can download the thesis memo below.

Investment Memo Adyen

417KB ∙ PDF file

Download

Download

Conclusion: Adyen is a buy right now.

Step 7: Bonus: Create a one-page summary

The memo is not that long, but if you want a single page with only the signal, use this prompt:

“Create memo/04_one_pager.pdf: a single page with ticker, current price, base-case fair value, upside %, 3 bullets bull case, 3 bullets bear case, and the valuation summary table from the DCF tab. Clean, skimmable design.”

You’ll get something like this:

Step 8: Track the investment

So, we have a single photo for Adyen today. But what if new earnings come out?

You just set up a scheduled task within Cowork.

“Create a scheduled task called ‘ADYEY Update’ that runs every Monday at 7am. On each run it should: (1) pull fresh price data via yfinance and update the historicals file, (2) check adyen.com/investor-relations for any new filings since the last run and download them, (3) if a new earnings release is found, update the model with the new quarter and write a short memo/earnings_update_YYYY_MM_DD.docx comparing results to my thesis in memo/03_thesis.docx, and (4) email me a one-paragraph summary of what changed.”

Now, bear in mind that your computer needs to be on and Claude needs to be activated. This is not a task that will work all the time. You could, for example, adjust the timing so it runs when you’re working on your laptop (maybe at 9 AM on a Monday).

(If you want tasks that run all the time, you need an always-on machine like a Mac mini, where your Claude Cowork is open. Or you can do it like me, and have a Mac mini where a script built with Claude Code does all of this every week.

Conclusion

These 8 steps provide you with an investment memo and a scheduled task for a single company.

You’ll need to decide for yourself if you’re happy with it. And iterate if needed.

What’s next?

Now comes the real power of Cowork and Claude. We did this once for Adyen. We don’t have to repeat all the steps for another company.

You can build a skill from this and run it at any company you want. Cowork will perform all the steps in this article one by one. You just need to point it to a directory where it can store and process the files.

Bonus step 9: Building the skill using the skill-creator skill

Yeah, it’s a bit meta. But it works. Claude has a skill to create skills. (We covered this before)

Just use this prompt:

“Use the skill-creator skill to turn my Adyen research workflow into a reusable skill called equity-research. The skill should take a ticker symbol as input and reproduce everything I just did for ADYEY: create the folder structure, pull yfinance data, build the historicals Excel, draft the business overview, key questions, DCF and comps model, thesis, and one-pager. Use the files in ADYEY/ as the reference example of what the output should look like. Keep the DCF assumptions parameterizable so I can adjust them per company.”

And it will create an equity research skills file in your folder.

Now go to Customize > Skills and add the file to your skill list. Once you do, you’ll see this:

Now, as a test, I ran the skill on another company, for example, IREN.

But before blindly executing the skill, always ask Claude the plan first:

Use the equity research skill to analyze a company called IREN. Show me the plan first before executing

The advantage is that before it actually starts using the skill, you can tweak it a bit.

Now, if you know nothing about this company, it will be difficult.

In this case, let’s just run it as a default skill run and see what it does.

Once Claude is running the skill, you’ll see how it advances through all the steps on the right side of your screen:

Note: I made a small modification to this skill to fetch annual reports from EDGAR, since this company is US-listed.

SO it took a couple of minutes, and now I have this:

With a memo and valuation model for IREN.

It considers IREN to be fairly valued, with a wide variation in possible future outcomes. That makes sense.

Final Conclusion

The bottom line is you can build things fast. And compound on the things you build.

The real work is the iteration and refinement to achieve the quality you need.

But once you’re past that hurdle, you can automate everything.

But always be mindful.

Claude is excellent at using your frameworks and your assumptions to give you a clear picture of a company.

But do not let it judge. You’re the judge.

Keep learning AI. Keep building.

Kevin

But which one is the best for stock analysis?

And can we build something ourselves that is even better?

To answer these questions, we first need to understand how LLMs “search”.

How AI search works for stock research

All these tools search in a specific way.

Here are the main steps outlined:

1. You ask a question/prompt: Find all information on company X. Only use official company information. No third-party websites.

2. The LLM will not add your prompt to its search API. It will decompose the prompt into several search queries

3. Now it will use its search API, mainly Bing and Google (just like we would)

4. It then scrapes the TOP 10 to 20 URLs for each query and cleans the text

5. Now it needs to divide and sort the contents for relevance

6. The LLM will synthesize everything and answer with citations

Now, in the flowchart, you can see I added something called Council Mode.

We’ll get back to that later on.

So whether you’re using Claude, Gemini, or ChatGPT, it always works the same way.

A Google search on steroids.

Which search engine do AI models use? (Google vs Bing vs Brave)

Here’s what the main LLM’s are using:

Nobody except Google is using Google Search. Almost everyone is using Bing. Only Claude is using Brave.

Subscribe now

The best AI search tool for equity analysts

When it comes to stock analysis, we want 3 things:

• Maximum amount of data

• Correct citations of the data

• No hallucinations

There are some research papers online in which researchers tested and compared different models.

Let’s first go to an older paper.

In June 2025, Chinese researchers conducted a comprehensive analysis.

The final result was the following:

Where different metrics were tested.

Both Perplexity and Gemini stand out, where Perplexity shows the highest accuracy in citations, and Gemini shows the highest effective number of citations.

But this analysis does not include the newest models.

The most recent paper dates from February 2026.

Here, Perplexity, combined with Opus 4.6, is the winner.

And more specifically, when it comes to specific Finance tasks:

Perplexity with Opus for Finance seems to be the best performer.

Now, let’s take a final, deeper look at the criteria we outlined: Accuracy, Depth of research, and citation quality.

Perplexity takes the lead once more.

Be mindful that it does not score 100%!

Now, although that was an independently reviewed paper, there is only one problem.

The paper was written by Perplexity.

And there is one important aspect we have not yet highlighted. Council Mode.

Multi-Agent “Council Mode”: why it wins

Imagine you analyze a company. And after analyzing, you discuss your ideas with other investors who have also analyzed the company. Often, it will shift your perspective about the company a little bit.

Your research has become stronger.

The final analysis is better.

What works for us humans works for LLMs.

But let’s first check whether this truly matters. And the best way is to look at what happened with Grok 4.2

Grok 4.2’s Multi-agent approach

This model was released with a special beta feature:

• You add your prompt

• Grok 4.2 will assess each prompt by 4 different agents

I tested it by copy-pasting one of my latest stock pitches and seeing what it has to say about it. Here’s what happens:

After 44 seconds (rather fast), it gave me this:

And Benjamin was the one who explicitly looked for weaknesses in my pitch:

Their overall conclusion: 4.5/5, great pitch.

Now, what matters more is that, because of this multi-agent mode, hallucination rates for Grok 4.2 have sprinted past Gemini 3 Pro and Claude Opus 4.6

That means multi-agentic analysis works.

Note: You need a Supergrok subscription to get this agentic analysis. ($30/month)

Perplexity Council mode for stock research

Perplexity was one of the first to offer a council mode.

In their max plan, when you prompt perplexity, it runs it through different models and reports where they agree and where the key differences are.

Don’t underestimate the value of this. Again, it increases the accuracy of the final analysis.

Note: Only available in Perplexity Max ($200/month)

But for those who don’t want to cough up $200/month, and are tech-savvy, here’s a “free” solution:

Multi-agent equity research

This GitHub project lists the code you need to build a multi-agent project.

With over 40,000 stars, it is popular. And it goes beyond what Grok and Perplexity do.

But it is more suited for traders:

It parses fundamental, technical, and real-time data before giving buy and sell evidence to the trader.

Building a custom researcher

In our previous articles, you’ve seen how we built Hunter to evaluate a large number of companies on a specific framework. You can find the results of the 95 software companies we analyzed here.

We also built a tracking system that, every week, gives us information on whether our thesis for our portfolio companies still holds.

Our next step is to build the researcher.

The researcher will use financial data, filings, and the latest news. It will complement that with a deep research function and output a full research report on the company.

As you can see in the image, we have a separate “red team” that does the same thing but takes the opposite stance.

But that’s food for a future article.

My recommendation: The best setup for stock analysts

When it comes to research, Perplexity and Gemini deep research score the best. Grok 4.2, thanks to its multi-agent mode, now significantly reduces hallucinations.

But if you want something customized to your needs, you’ll need to build it yourself or adapt an existing project.

Keep learning AI. Keep on building.

Kevin

Share

Your typical screener uses financial criteria, or you could even just type what you want as a prompt these days.

But now you’re leaving money on the table.

You’re not really using the power of these LLMs.

What we’ll show you today is a tool we built called Hunter. Because of the massive drawdown at many software companies due to the rise of AI and the development of tools like Claude Code or Claude Cowork, I set out to build a specialized software screener to analyze these companies and assess how AI-resilient they really are.

The software screener

It looks like this:

The inputs are 10-Ks and 10-Qs to provide context on the different companies and financial data. The brain is a Python pipeline that fetches data and uses Claude Sonnet to apply a scoring framework to the companies. The final output is an Excel or CSV file with a list of companies.

The AI resilience scoring output looks like this:

And each score has an analysis behind it.

Here’s a glimpse of Duolingo:

Now, the result I’ve shown here was an intermediate result while we were iterating. The final result is a lot better, thanks to Einstein’s help.

Everything should be made as simple as possible, but not simpler

- Albert Einstein

In the past, you could only screen on financial data.

Now, we can screen on text. But as we said before, AI is bad at judgment. We need to help it. That’s why you need a framework. Some kind of scoring system.

The cooking recipe

This build is like cooking. You need ingredients and a recipe. But first, what did we use?

I built this in Claude Code. And I wish I could build it in Claude Cowork, but that’s not possible. I will be able to provide it as a tool you can use in Claude Cowork.

The ingredients

• A scoring framework. We are using a specific AI resilience framework to score software companies. But you can imagine other use cases. For example, you could build a framework on serial acquirers or on turnarounds. Be as detailed and specific as possible, and make sure the LLM can score a company based on text. The scoring framework is stored as a markdown file.

• Financial data. If you don’t have a database API, you can always use yfinance. It’s an unofficial financial data API that scrapes data from Yahoo Finance. It’s great for personal use, but be mindful that it can always stop working tomorrow.

• A list of tickers: You need a list of companies you want to evaluate. I pulled them from my financial database. You can always just pull a couple of companies from your portfolio and add them.

The recipe

Now we need to bring all the ingredients together in a certain way. Our output is an Excel file that scores each company and lists some financial data.

The workflow looks like this:

1. Ingestion layer: Retrieve the data from the EDGAR filings and the financial database

2. Cleaning layer: Process the data so it’s more suitable for an LLM to process

3. Processing layer: Parse the data and run the scoring framework

4. Output layer: Output a CSV and Excel file

One thing to note is that this requires iteration. When you build this, it won’t be perfect from the start.

And I encourage you to just start building something small like this, testing it on a couple of companies, and keep building. You’ll be surprised and amazed by the final outcome.

Here’s the simplified version:

Subscribe now

Using Claude Code

So how can you build this yourself?

When you use Claude Code, always start with your end goal. What do you want to achieve? Be as precise as possible.

I asked Claude the following:

I want to build a tool that uses financial and text data to analyze software companies, score them on AI resilience, and output CSV and Excel files with a row for each company, including its score and selected financial metrics.

And then ask Claude:

Ask me questions? What additional context do you need to achieve this goal?

Claude will start the interview. The goal is to develop a plan. Once you and Claude are happy with the level of detail in the discussion, you say:

Take all the information from our brainstorming sessions and show me a step-by-step plan that I can follow to build this.

Claude will go into planning mode and show you the plan.

Review the plan, make small modifications if needed, and then just start building from step 1.

Now there might be 20 steps. And once you start to test, you’ll run into edge cases that need solving. I’m not saying this is a breeze. But you can unlock dev powers for yourself.

I’m not a software developer. I’m not really that tech-savvy.

I consider myself a power user with endless curiosity.

If I can build this, you can too.

I’ll see if I can add a detailed video on our course to show you how to build this from start to finish.

Next steps

We’ll run Hunter through about 100 to 200 US Software companies, and I’ll provide you with the final results.

And my aim is for Hunter to become more flexible. To allow you to select frameworks and use them on certain companies, and see how they score.

If you want to learn Claude Code, how to build this, or just download the entire project and run it on your own computer, you can join the private beta for our course and skill library with the button below.

JOIN PRIVATE BETA

Note: I want to emphasize that the private beta is not free. You’ll get access to a platform with prompts, skills, videos, and a software tool called Flow to track your decision-making. You’ll be able to enter at a significant discount off the final release price, but in return, I will need your feedback while you're “playing” with the product.

Thanks again, and keep learning AI, keep building.

Kevin

Claude is constantly being expanded.

First, we had a simple Claude Chat; then they introduced Claude Code and Projects; and a month ago, they launched Claude Cowork.

And at the end of this series, we will show you how to build an automatic thesis tracker that will scrape data from the web every week and show you if your thesis is still intact.

Today, we’ll show you how to use Claude projects to build an investment memo. However, the method described below can also be used to generate earnings reports and related outputs.

Introduction: How Claude Projects work

You might have gotten used to chatting with chatbots. You craft a great prompt, and then keep building until you get your desired result.

And an hour later, you have another idea, but your chat window is full. You’ll need to start again. That new chat has no context for what happened before. (or limited with Claude’s memory)

One solution to this is to use Claude Projects:

The goal is to have one location where all your context is stored. So each time you start a new chat, you don’t start from 0.

You have the possibility to:

• Give it a system prompt: This is like a god prompt that Claude will always read before it starts answering you in chat

• Give it a knowledge base: these are documents, data, or whatever you want to feed it to provide additional context.

And as we said previously, it’s a genie in a bottle, but you need to constrain it.

So let’s apply it to a company called IQE PLC, a hot company in the photonics bottleneck craze for datacenters.

Here’s a view of the project we created:

So let’s go through the steps on how to set this up.

Step 1: Define a goal and a system prompt

In our example, the goal is: IQE Investment analysis. This means, within this project, we can generate an investment memo, create an earnings report when a new report is available, and build a valuation model, all within the same project.

The system prompt is a general prompt that is always “active”. It’s up to you to choose this, but here’s a simple one you can always use:

You are an investment research analyst helping me build and maintain
 an investment thesis for a specific company.

Analyze financial data, earnings reports, competitive positioning, and industry trends
Help me identify key drivers of value (growth catalysts, moat, management quality,
 capital allocation)

Challenge my assumptions with bear-case arguments when I present bullish views, and vice versa

Flag risks I might be overlooking

How to respond:
Start with the bottom line, then provide supporting detail
Use plain language, avoid unnecessary jargon, and explain it when you do use it
When referencing numbers, always note the time period and source
Distinguish clearly between facts, estimates, and your reasoning

Thesis structure to maintain:
Follow the thesis.md file.

Important rules:
If I upload earnings transcripts or filings, extract the most important data points, and flag anything that changes the thesis
When I ask you to update the thesis, show what changed and why

It’s up to you to refine and improve upon this prompt. You can make it more or less focused; it depends on the project’s goal.

Step 2: Upload your data

Now that the project is defined, you will need to feed it data.

Claude gives you 4 options:

• Upload files from your computer (that’s what we did)

• Add text content, but type it in directly

• Connect to GitHub so that it can read files or code from there

• Or read files from your Google Drive

This last option can be useful if you store all your annual reports or earnings updates in Google Drive; you can grab them from there.

We retrieved the last 2 annual reports and half-year reports from the company and uploaded them manually.

Then we added a thesis template as a Markdown file. This is optional, but we want the final memo to follow a very specific structure. You can do the same for an earnings update, for example.

Your project is now ready. You can start chatting with it.

Subscribe now

Step 3: Start analyzing

Now that the project has a system prompt and the data, you can start your analysis.

Open a chat and write a prompt that generates the investment memo. You may have an earnings update template, or you want to generate a full research report. This is all possible inside the project.

And if you go back to the top, you’ll see that I’ve selected Opus 4.6 extended thinking.

In this specific use case, because you’re asking it to generate a report, use the best version, which means the best model, with the longest thinking time.

(Use cases where this is not needed: If you’re just looking for a quick back-and-forth chat, Sonnet might be enough. Or if you just want to summarize a document, Sonnet might also be enough.

Pro Tip: If you generate a report, upload it to the files, so that the overall context window grows. And iterate on that context in the next chat. Claude will not load all the files into the chat. It uses retrieval augmented generation, it indexes the content, and it will only retrieve the data that is relevant to your prompt.

In that way, you can create a living memory.

And remember, Claude recently rolled out a memory function, so it is able to retrieve information from previous chats. Go to settings and capabilities and check if you activated it or not.

So let’s now build the thesis for IQE:

You can download the full investment case below.

Iqe Investment Thesis

158KB ∙ PDF file

Download

Download

Pro Tip: If you can, add financial data as an Excel directly into the projects. This avoids Claude having to extract the data from a PDF.

That’s it, a couple of minutes of work. An investment thesis like this can be a great starting point to launch your research.

Keep learning AI, keep building!

Kevin

If you’re interested to learn more, download prompts and skill files, and have your own stock research engine.

We’re launching a private beta of our software tool soon. You can use the link below to join the waitlist.

JOIN WAITLIST

Note: Private beta will be paid. The goal is to iterate with the users to create a better product. Only for serious investors.

Then you have to use Claude’s skills.

It will create leverage.

All screenshots taken come from the Claude Desktop app.

What are Claude skills?

It’s basically a prerecorded instruction that you can invoke when you wish. Like a small program, designed for llm usage, that can create stuff for you.

Here’s what Antropic says:

Skills are folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks. Skills teach Claude how to complete specific tasks in a repeatable way, whether that’s creating documents with your company’s brand guidelines, analyzing data using your organization’s specific workflows, or automating personal tasks.

You can find it through the customization button.

Customizing means:

• You can connect other programs

• You can create new skills

And if you go to create new skills:

You’ll see your own skills (this will be empty if you start with this)

And you’ll see examples of premade skills you could use.

Claude has a skill-creator skill to create skills.

Very meta.

Example of a custom skill

I created a thesis alignment skill.

Every quarter, I will use this skill to generate an earnings report that verifies if the original thesis still holds.

I can use this skill anywhere, in Claude Chat, Claude Cowork, or Claude Code.

Here’s a video applying this skill to Rubrik (RBRK).

In this case, it will generate an earnings report and judge that, based on the last quarter, Rubrik was overperforming.

Subscribe now

How to build a skill?

Simple, but not easy.

Here are the steps:

1. Ask Claude to help you

2. Let Claude search for information on the skill you want

3. Think about the output you want: Just text, a file, a website?

4. Now, based on all this information, ask Claude to build this skill with his skill creator skill

5. Iterate until the outcome is to your desire

You ask Claude to help you. With a prompt like this:

You are a master in skill creation. I want to create a skill to [enter your goal]. 
Give me a plan on how to set this up. 
Use your creator skill as a reference to give me the necessary instructions.

And Claude will give you a plan that you can then follow. Then follow the remaining steps.

I say it’s simple but not easy, because in my experience, there is a lot of iteration involved to get to a high-quality outcome.

But that’s the cool thing. Claude will give you a markdown (.md file), which is the skill you wanted. If you want to make modifications, just chat with Claude, and once you know what you want changed, ask Claude:

Please make these changes … to the skill markdown file you just created.

And you keep iterating.

Once you’ve built a skill, Claude usually will give you the option to add it to your skill library with one click. But if that option is not available, just go to Customize, skills, and click on the + button.

Upload your new skill:

Now, as you can see from the image, skills can range from simple to more advanced:

• The earnings report generator is a single markdown file

But I have other skills that have multiple markdown files, and as you see for the creator skills, it’s a mixture of markdown files, HTML files, and Python scripts.

You can imagine a valuation skill that holds a Python script and generates an Excel with a valuation model, based exactly on your own models.

Now, Anthropic is building these kinds of skills itself. Unfortunately, I don’t have access yet. I’m still on the waitlist.

If you want to start building these yourself, the sky is the limit.

But I want to give 2 recommendations:

• Start small

• Focus

In my experience, it’s better to have several smaller skills that produce high-quality output as opposed to one huge skill that creates a 40-page detailed company analysis.

Some ideas are:

• Earnings call transcript analyzer: Upload a transcript, get key takeaways, management tone shifts, and forward guidance extracted automatically

• Competitor snapshot: Given a company, auto-research its top competitors and build a quick comparison table

• DCF model builder: Input a few assumptions and get a discounted cash flow valuation in a spreadsheet

• Risk factor analyzer: Condense long risk sections from filings into a ranked, plain-English list

• Earnings calendar prep: Given a list of tickers, build a pre-earnings cheat sheet with key metrics to watch

• Portfolio exposure analyzer: Upload your holdings and get sector, geography, and factor exposure breakdowns

If you have a great idea, you can build it.

Institutional grade equity reports

For example, this Claude equity research skill has been built specifically for Claude Code. You won’t be able to simply upload it into Claude Chat or Claude Cowork.

Now that’s a bummer. I know how to use Claude Code, but most people don’t.

But there’s a simple solution.

Ask Claude to read the GitHub and create a skill from it.

This took Claude less than a minute.

And now I have a new skill in my library: Equity research skill.

I applied it to Rubrik, and you can see what it generated here.

Honestly, that’s a pretty decent report!

Bottom line, the sky is only limited by your own imagination.

Keep learning AI. Keep building!

Kevin

If you want access to our earnings report generator skill and other skills I’ve built, you can sign up for our private beta.

JOIN WAITLIST

That private beta is a first version of a prompt/skill library complemented with video instructions and a lot more. The private beta will be offered at a significant discount, but in return, I ask for feedback to improve everything.

In a future article, we’ll go deeper into connectors and how to connect other programs to Claude.

I’m not talking about your financial software, where you have your portfolio, watchlist, and financial metrics.

Probably any software package can give you that.

I’m talking about something that is customized to you. With your context. With your prompts. With your data. And especially with your investment cases.

The future is not building new software. The future is building custom software that fits your exact needs.

The main advantage I see from a daily brief is to track thesis drift.

Based on new information coming in, does my original thesis for the company still hold?

Remember what we learned from Richard Feynman:

• You postulate a hypothesis (thesis)

• You can explain that thesis in simple terms

• You always try to test and break the thesis

Based on this, as long as the thesis holds, you hold the company.

And here’s a sample of what it could look like:

Or here’s a video if you want more information on the full dashboard:

Always start with the end in mind. So the image and video are the final build.

But it has lots of moving parts.

• Sourcing new information on all the companies

• Adding it to a database for each company

• Generate the current state of the company

• Compare it to the initial thesis

• Provide a conclusion: On track, Watch closely, or Broken

• And lots of other information like earnings calendar, important upcoming events, etc…

First principles tell us to break the problem into its tiniest parts. Then solve each problem, and build back up again.

So today, I’ll show you a simple way to source and analyze new information from a company you like or own.

Subscribe now

Before jumping into ChatGPT, Claude, or any of the models out there. Ask yourself this question:

• Can I get the same result more simply?

• Do I really need an LLM to do this for me?

Why do I ask this?

Because you can quickly end up spending hours setting things up. AI Chatbots are fun to use, and you can get trapped in testing stuff all day long.

The time opportunity cost is real.

For example, if you want the most recent updates from a specific company:

• See if you can subscribe to their email list

• Maybe the company has an RSS feed

• Consult third-party websites that do this for you (For example, I use

https://investorpa.com/

• for companies in Australia)

Now, when it comes to creating a daily brief, there are 3 levels:

1. Use ChatGPT tasks

2. Use Zapier to automate

3. Use Python to build a script

Now, let me show how to build the simplest daily brief of all:

• Fetch the latest news for one specific company (daily, weekly, as you wish)

• Receive an email with that news

Level 1: Use ChatGPT tasks

Surprisingly, ChatGPT is the only one that can handle automated tasks natively. Maybe Claude Cowork can, but I haven’t tested it yet.

You open a new Chat in ChatGPT and simply ask it to retrieve information on a company on a daily basis. Like this:

Once a week, I receive an email in my inbox with the latest news.

Like this:

The pros:

• Simple to set up

• You can do this for multiple companies

• Cheapest: You only need a $20/month ChatGPT subscription

The cons:

• Over time, the context window will get crowded

• No real data storage can lead to repeated messages

Now this is an extremely simple example, but if you use a more detailed prompt like this, you will get something that resembles an actual brief.

Play with it so that it suits your needs.

Prompt:

Every week, on Monday at 9:00, run a task and send me an email with the following structure: 1. 3 market events of the week. Screen for the most important events in equity markets over the past week, provide them in a bullet point format, and add the URLs so that I can read further if I want. 2. News on portfolio companies. List all the new information available this week on the following companies: TSLA, NVDA, and MSFT. Provide a table with multiple columns: Ticker, Data, Summary News, Url 3. News on watchlist companies: Do the same thing, but for companies on my watchlist. RBRK, NUAI, ADYEN Run this task now and show me the result

What happens is that ChatGPT will run that one prompt every week and add the results to that same chat window. You can imagine that over time, the chat window will become lengthy.

So if this is the setup you’re using, after a certain time, you’ll have to rerun the task in a new chat.

ChatGPT will use its web search tool to fetch the data, and then use its LLM to analyze it based on your instructions. It will then send you an email to alert the task has been completed, and you can check the results.

Level 2: Automate with Zapier

The goal of ChatGPT is to use its powerful LLM to analyze text and data. But using its task feature can be limiting.

We also want more control over the data retrieval.

So here’s the tech stack you can use:

As you can see, all tasks are now separated.

The data feed will be handled by RSS.app. RSS is a structured way that websites can use to feed information (like a new earnings report). The RSS.app allows you to create a bundle. So let’s say you have 10 companies with their RSS feed. You bundle everything into RSS.app.

So now you have the data, but you need to feed it to ChatGPT. The easiest non-code way is to use Zapier.

And before you think, heck, Kevin, do I need to learn Zapier too?

Let me show you how easy this is.

You don’t need to be a Zapier specialist. You just need to be curious and use Claude or ChatGPT to your advantage.

I built this in 5 minutes:

It holds all the steps as previously discussed, and it sends me an email every day.

The trick to setting this up is the following:

1. Ask Claude or ChatGPT what you want to build in Zapier. Iterate a little bit. Explain in as much detail as possible what you want to achieve.

2. Now, since Zapier uses Copilot, ask ChatGPT to give you the exact prompt you need to put into Zapier.

3. No,w copy paste that prompt into Zapier, and Zapier will build the above sequence for you

The only thing that’s left is that you’ll see windows opening up to ask you to:

• Copy-paste the RSS url in the first step

• Copy-paste your ChatGPT API Key

• In the Gmail step, connect with your existing Gmail account

Now, why go through all this trouble?

Well, remember, we want to build that fancy dashboard

And instead of letting it send you an email, you can ask it to send the information to a Notion database or a Google Sheet.

Now you have data in the background to populate the dashboard.

Level 3: Python script

If you don’t want to use something like Zapier (or Make) for automation, then there is only one option left.

Write a script in Python that runs the automation.

Again, it’s not as complicated as you think if you know how to leverage Claude Code.

But we’ll cover that in a future post.

Final setup

But do you really need all of this?

I use it in 2 ways:

For each company, I run a weekly brief. I read that brief to see what’s happening. (I don’t need a daily brief, I’m a long-term investor, not a trader).

For each company, I have a folder. That folder holds:

• 5 years of annual and quarterly reports

• My initial written thesis

The weekly briefs are added to that folder every week. Once a quarter, a new quarterly report is added to that folder

So the goal is to have maximum context in one single folder. Like a living memory.

Now I run a quarterly thesis alignment skill on that folder. It generates a report that shows if, with all the current information, my thesis for the investment case still holds up. It also includes an earnings report based on the structure I want.

Then the thesis alignment report is added to the folder as well.

As such, it keeps building its memory over time.

You set this up once to reduce the time spent tracking the company.

The process looks like this, and the result is added to the dashboard I showed before:

If you’re interested in the whole system, a private beta of the website we built that shows all of this will be launching soon.

The private beta will not be free. It will offer the prompt/skill library and video course at a significant discount. In return, I ask for feedback on the website itself so that we can improve it.

It’s only for serious investors who want to spend time learning all of this. Spots will be limited.

You can join the waitlist for the private beta here.

JOIN WAITLIST

Keep learning AI. Keep on building.

Kevin

Share

While trying to build the atomic bomb, there was only one schedule that mattered:

As fast as possible.

He needed that note to continue his work, and the person with the key was nowhere to be found

So he did what problem-solvers do: he walked to the row of these “secure” filing cabinets and treated the locks like a puzzle. In an open drawer, he noticed a sheet with the Greek alphabet, and across the top someone had written “π = 3.14159.” He tried combinations based on π. No luck.

Then he made the next, more human guess: if someone chose π as a hint, they might have chosen another scientific constant. So he tried the constant e = 2.71828 and dialed 27-18-28. The cabinet opened.

So later on, he started similarly picking the other locked cabinets and left notes in each of them.

The note said: “I borrowed document xyz - Feynman, the safecracker.”

His name was Richard Feynman, who in 1965 won the Nobel Prize for physics.

Richard was a tester, a tinkerer. He didn’t take anything for granted.

“The first principle is that you must not fool yourself—and you are the easiest person to fool.”

How Richard Feynman's first-principles thinking applies to stock analysis

If there’s one area where this fallacy holds more than any other field, it’s stock investing.

Most investors, when they think they found a great investment case, will go look for confirmation bias in investing.

• Who else is buying?

• Who else has written about the stock?

The better approach would be to continuously try to break your thesis, try to break into your own safe.

Be a safecracker.

That’s where the Feynman Technique comes in:

• Set the hypothesis

• Test it until it breaks

Let’s go a little deeper into how it works.

The Feynman Technique is a structured way to stress test your thesis (or anyone’s thesis).

There are 3 big actions to take:

1. You try to explain the business in plain language, like you’re teaching it.

2. Every time you get vague (“pricing power,” “platform,” “network effects,” “AI advantage”), you mark it as a gap.

3. You replace each gap with something testable: a mechanism, a metric, and a disproof test.

Then you repeat the process over time, because new data is constantly trying to tell you your “combination” is wrong.

That’s it.

It looks simple, but it’s actually hard. It’s a first principles approach to investing.

Luckily, we now have LLMs to help us do this. (and an AI-enhanced stock analysis process)

So let’s build a prompt using the Feynman Technique.

Building an AI prompt using the Feynman technique

Here’s how to go forward:

The refinement phase is important. Don’t just expect the LLM to give you the final answer. You need to tweak it.

The goal of this prompt is to push back on your thesis. To identify gaps and provide measures to track them.

The Feynman technique shines on a very specific and focused question.

Now the prompt:

# Persona
 You are a Feynman-style business explainer and adversarial investor. You distrust buzzwords, separate facts from claims, and treat every thesis as “not yet falsified.”
# Resources
 Use only: The bullish pitch I paste. Official company-first sources only: SEC filings (10-K/10-Q/8-K/S-1) or other official filings, annual reports, shareholder letters, investor presentations, earnings call transcripts, official IR press releases, and the company’s website/pricing/docs. Generally known accounting/finance logic. If required data isn’t in these sources, say so explicitly and list what’s missing and where it would normally be found; do not invent numbers.
# Instructions
 Turn the bullish pitch into a thesis test harness to monitor each quarter. Prioritize falsification: identify what would prove the thesis wrong early. Avoid jargon; if unavoidable, define it in plain language. Be blunt but not rude.
# Method
Feynman explanation: 8–12 bullets: Translate the bullish pitch into a 12-year-old explanation: who pays, for what, why now, why repeatedly, how value is delivered, why customers don’t switch, what drives growth and costs.

Extract claims/gaps: Rewrite every vague statement as a testable claim (e.g., “pricing power” → “ARPU rises without churn rising”). Flag missing info.

Mechanism + disproof test: For each claim: mechanism (cause→effect), strongest supporting evidence from official sources, disproof test, leading vs lagging indicators.

Build the scoreboard: 10–15 thesis tests max. For each: primary KPI, supporting KPIs, expected range/trend, red-flag thresholds, source location (filing section / transcript), and 1–2 anti-Goodhart cross-checks (how the KPI can mislead + what to track alongside).

Failure chain: Top 5 failure modes + earliest signals.

#Expression (Output)
 A) Feynman's explanation
 B) Claims/gaps (testable statements)
 C) Thesis Test Harness (per-claim KPI set + thresholds + sources + cross-checks) -> provide a table with a column that assesses if the red flag is currently active or not
 D) Missing info to verify next
 E) Failure modes + early warnings

So let’s test it on my Adyen investment pitch, I wrote (Ticker: Adyen).

Subscribe now

Putting the Feynman prompt to the test: Adyen case study

I’m not going to copy/paste the entire output, but here’s a table it made. What’s interesting is the red flags column. Based on your knowledge of the company, you should refine this column and then ask your LLM to assess the company based on its latest earnings reports to see if any flags become active.

Here’s the final result:

Which means from a business perspective, no red flags have been raised at this time. (although the market has discounted the company)

When all is said and done, what’s important is that you can explain your investment thesis in simple terms without making it too simple. Use this prompt to push back on your thinking and strengthen your thesis.

Keep using AI. Keep on learning!

Kevin

P.S. We’re building a library of prompts, Claude skills, custom GPTs, and more. In a couple of weeks, we’ll launch a private beta.

This means you have the opportunity to test a working product that still needs refinement. You’ll be able to buy it at a significant discount, but in return, you’ll need to complete a survey and offer feedback to improve the product.

If you’re interested in this, just click the link below.

JOIN PRIVATE BETA

They’ve been playing leapfrog since 2022:

• 2023 and 2024 were the years of ChatGPT, releasing several models and nailing the benchmarks (even though Grok also released in 2024)

• 2025 was the year ChatGPT’s dominance started to waver and the rise of Google and Anthropic:

  • March: Gemini 2.5 pro launches

  • August: GPT-5 launches, reclaiming the benchmarks

  • September: Claude Sonnet 4.5 is claimed to be the best coding model

  • November: Gemini 3 pro takes the lead

But does it really make that big of a difference in your choice of LLM?

Today, we’re going to take a look at:

• Which models are used the most?

• Which models are the overall best?

• Which model is best suited as a finance agent?

And more importantly, which ones you should absolutely avoid!

But first, which model are you using?

Which LLM do most people use?

Measurement method: Similarweb estimates websites’ traffic based on openweb data for certain websites. It does not include API usage. It does not include embedded usage (like Microsoft Copilot or Gemini within Google workspaces).

The results:

From a traffic perspective, Gemini is taking market share from OpenAI. All the other models seem to be playing around in the margin (Anthropic sits at 1.8%). But as we’ll see later, that doesn’t mean you should ignore them, though.

What are the developers using?

Measurement method: Number of tokens consumed by the model over OpenRouter. OpenRouter is open source, so enterprises will refrain from using it. Don’t take the results as: This is the current API winner!

A developer will use a model that excels at coding and will also take API costs into account (the lower the better).

This is data from OpenRouter. A big shift has occurred with the release of OpenClaw (see our initial article on it). Especially, the use of the Gemini and OpenAI API has reduced in favor of Moonshotai with their new Kimi K2.5 model. (It’s a Chinese open source model)

You can also see the wide usage of xAI at a certain point in time.

But this is very niche, let’s see what companies are doing ⬇️

What are the enterprises using?

Measurement method: A survey on model usage across 150 startup and enterprise leaders performed by Menlo Ventures.

The impact on OpenAI is stark here. Anthropic is the clear winner, and with the release of Claude Cowork and Claude for Excel, they are going after the enterprise client. The results are probably skewed, though, as I believe this specificially show which LLM is used for coding.

The big competitor will be Microsoft. If Microsoft can implement AI in a better way in its software suite, then Claude might be in for a difficult ride.

Now let’s avert our attention from the usage to the actual benchmarks.

The benchmarks

A benchmark is usually a list of questions that is presented to the LLM, and the answers are graded. Be mindful that these can be gamed just like a human can game an exam.

We’ll look at this 2 ways. Which LLM is the best for general purposes, and which LLM is the best for financial analysis?

For general purposes

We’ll use Omniscience by artificial analysis. Here’s what they do:

AA-Omniscience is a knowledge and hallucination benchmark that rewards accuracy, punishes bad guesses and provides a comprehensive view of which models produce factually reliable outputs across different domains.

The benchmark contains 6,000 questions across 6 major domains, derived from authoritative academic and industry sources and generated automatically using an LLM-based question generation agent to ensure unambiguity, scalability and factual precision

AA-Omniscience Index (higher is better) measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer. Scores range from -100 to 100, where 0 means as many correct as incorrect answers, and negative scores mean more incorrect than correct.

Here are the current results:

So that means the best model, which seems to be Gemini 3 pro, scores a whopping 13/100.

Grok 4.1 fast score -31/100. What’s surprising is that even the Gemini 3 flash model has a pretty high score.

Note: There are no special techniques used to reduce hallucinations. This is just a question-and-answer round.

Subscribe now

For financial analysis

Vals.Ai, in collaboration with researchers from Stanford, investment banks, and industry experts, has created a list of 537 questions to evaluate the financial skills of an LLM. The goal is to see how close the performance of an LLM comes to that of a junior analyst.

Here are the most recent results:

Claude Opus 4.6 jumps to the top spot, with a significant improvement over the previous models. That’s not surprising, as Anthropic wanted to focus more on finance and research knowledge work.

But 60% seems low?

Well, it increases for specific tasks. Here’s an overview of specific financial tasks and how the 3 most recent models rank:

Claude Opus takes the lead, followed by GPT5.2. Gemini performs less on all metrics except for Financial Modeling projections.

If you want to know what each task means, you’ll find some examples below:

Quantitative Retrieval (easy)

Direct extraction of numerical information from one or more documents without any post-retrieval calculation or manipulation.

What was the quarterly revenue of Salesforce (NYSE:CRM) for the quarter ended December 31, 2024?

Qualitative Retrieval (easy)

Direct quotation or summarization of non-numerical information from one or more documents.

Describe the product offerings and business model of Microsoft (NASDAQ:MSFT)?

Numerical Reasoning (easy)

Calculations or aggregation of key numbers to produce an answer.

What is % of revenue derived from AWS in each year and the 3 year CAGR from 2021-2024 of Amazon?

Complex Retrieval (medium)

Numerical or non-numerical retrieval or content summarization requiring synthesis of information from multiple documents.

Please briefly summarize the most recent capital raise conducted by Viking Therapeutics (NASDAQ:VKTX).

Adjustments (medium)

Quantitative and qualitative analysis of reporting context bridging GAAP and Non-GAAP Financial Metrics.

What is Lemonade Insurance’s Adjusted EBITDA for the year ended December 31, 2024?

Beat or Miss (medium)

Comparison of forward management guidance versus actuals, synthesized by reconciling sequential quarterly reporting documents.

How did Lam Research’s revenue compare to management projections (at midpoint) on a quarterly basis in 2024? Format as % BEAT or MISS. Use guidance provided on a quarterly basis.

Trends (hard)

Analyze patterns within a single company’s reporting structure or calculate and contextualize evolving performance, key metrics, or business composition.

Which Geographic Region has Airbnb (NASDAQ: ABNB) experienced the most revenue growth from 2022 to 2024?

Financial Modeling (hard)

Complex numerical reasoning calculations require additional financial expertise to define and evaluate.

How much M&A firepower does Amazon have as of FY2024 end including balance sheet cash, non-restricted cash and other short term investments, and up to 2x GAAP EBITDA leverage? Round to nearest billion

The bad performers

“The only thing I want to know is where I’m going to die so I never go there”

- Charlie Munger

Instead of searching for the best model, which clearly is Opus 4.6, in typical Munger style, it might be best to look for what we should avoid.

We found 3 things based on this analysis.

1. Avoid fast, instant. You want quality, you need to give the model time to think (Like a human would). Always use thinking mode.

2. Avoid small, mini. Most models come with variants, like GPT5-mini. With LLMs, size matters. Avoid small models if you want performance.

3. Avoid data retrieval by the model (feed it the data it needs)

A note on tool calls

A finance agent is often about retrieving sec filings and pulling information from them. This is not the base task of an LLM. These are tools calls, where the chatbot will invoke a specific tool to retrieve the information.

So the performance of the different LLMs is heavily influenced by how good it can repeat tool calls to gather the necessary information.

This also means that you can make these models a lot more performant by providing them with the data!

Key Takeaway

I’ve been watching Vals.AI benchmarks for over a year, and the biggest conclusion is not which is the best model, but that they keep getting better and better. And the pace of improvement is not stopping.

Here’s the overall finance agent performance in time:

This chart takes the best performer each month since 2024:

In the end, what matters more than picking Claude, GPT, or Gemini is how you use it.

Feed it quality data, and quality prompts, and you’ll get great results.

We’ll go deeper into the data and prompting in future articles.

To conclude: I have been using GPT5.2 extensively in the past, but after testing Claude Opus 4.6, there’s a new winner in the house.

What do you think? Is model choice overrated?

Leave a comment

Until soon!

IWAI

Let me know what you think of this article:

ChatGPT, Gemini, Grok or Claude.

They are genies.

Eager to grant your every wish.

But before going “full Aladdin” and making your 3 wishes, an LLM, while genie-like, has to remain in its bottle.

It needs to be constrained.

The bottle represents:

• The data you give it

• The context you add

An LLM does not think.

You think.

All these models act like genies once you tell them exactly what to do.

Avoid open-ended questions. Avoid asking about sentiment (without defining it precisely).

The more effort and quality you add, the more you will receive.

So, if we want to use it to act like Buffett, we need to tell it how to think like Buffett.

So how does Buffett read financial statements?

How Buffett reads financial statements

Buffett looks for signals that could indicate a long-term competitive advantage. But never forget the importance of price. I don’t think he ever paid more than a P/E of 15.

Let’s do a quick run-down of the 3 financial statements.

Income Statement

From revenue to operating profits

Gross Margins

➡️The gross margin (gross profit/revenue) needs to be high, ideally above 40%.

It could signal pricing power. It provides a cushion for the company when things go sour.

Next, he takes a look at how different operating expenses compare to gross profits between different companies.

He’s looking for companies that are drowning in cash, so the cost structure matters.

SG&A

Selling, General & Admin costs can eat away at profits.

➡️He looks for SG&A/gross profits lower than 30%. The lower the better.

R&D

Phil Fisher loved R&D because it is needed to fuel future growth. But a company that can sustain its future growth with low R&D spending might have something special.

High R&D costs could mean the advantage is temporary, that it is not inevitable, or structural.

➡️Look for R&D/Gross profits < 10%

Depreciation

Depreciation is a cost. You already know he despises EBITDA.

Even worse, the cash was already paid in full at the start of the investment!

Low depreciation costs could mean that it’s a capital-light business. And capital-light business might be more profitable.

➡️Look for Depreciation cost/gross profits < 10%

From operating profits to net profits

Interest expenses

You don’t need the balance sheet. The Income statement can give you a signal of financial health.

High MOAT companies can have debt, but their earnings power is so strong that if they need to, they can pay it back pretty quickly.

➡️Look for Net interest/Operating profit < 10%

Net Income and EPS

Buffett wants a highly profitable company. He wants the company to have a history of rising profitability, not only in terms of the amount but also in terms of earnings per share.

➡️Look for rising income and EPS

Income Before Tax and Normalization

There are no criteria for this, but I wanted to mention 2 important things:

• All investments are relative. Buffett likes the income before tax because if he divides it by the market cap, he gets an equity yield. This allows for easy comparison to a bond yield or another company.

• When he talks about earnings, he means normalized earnings. We need to remove one-off profits or losses, or even look at the cyclical nature of the business. Don’t just look at the P/E on a screen.

The 10-step checklist based on the income statement:

Subscribe now

Balance sheet

Let’s look at some important criteria from the balance sheet highlighted in red.

Cash

Buffett wants companies drowning in cash. So if cash and cash equivalents have been increasing over the last few years, it’s a good sign the company is doing something right.

➡️Look for a steady rise in cash and cash equivalents

Inventory

In the case of a company that holds inventory, a steady increase in inventory over the last few years signals the company is selling more and more products. This is a good sign.

Net receivables

The best companies have a cash conversion cycle that is negative. In other words, they get paid in cash first and only have to pay their own suppliers later.

➡️Buffett looks for net receivables/sales that are lower than competitors

Long-term debt

Debt in itself is not an issue. Strong companies usually have a low debt level. But even if they have debt, look at the earnings power compared to the debt.

➡️Buffett looks for a long-term debt-to-earnings power ratio below 4 (if profits are cash, they can pay back debt in 4 years)

Debt to equity

➡️Is debt-to-equity consistently low, preferably below 80%?

Retained earnings

This is one number that gives you information about the history of the company. How much earnings have been retained within the company in the past?

➡️Buffett looks for companies with a history of increasing retained earnings

Treasury shares

When a company buys back shares, it registers them as treasury shares. This reduced the number of shares outstanding of the company.

➡️Buffett looks for companies with increasing treasury shares.

Return on Equity

Has Return on Equity been stable or grown over the past decade? ROE is a measure of efficiency, although it might not be the best measure:

ROE can be a trap, as debt can inflate ROE. You might want to swap this for ROIC instead.

Cash Flow Statement

Profits are an opinion, cash is a fact

Alfred Rappaport

CAPEX

Capital investment is needed to maintain and grow a business in the future. But there are big differences in industries.

➡️Because Buffett is looking for a capital-light business, he wants CAPEX/Net Income to be below 25%.

Buybacks

We already covered this through the balance sheet, but the cash flow statement will also show the repurchase amount of shares (if any).

The 10-step checklist is based on the balance sheet and cash flow statement

Now let’s make a prompt out of this.

The prompt

Here’s an image of running this prompt on UnitedHealth Group (UNH) in ChatGPT. You can use it in 2 ways:

• Look for red flags: This allows you to screen through companies fast.

• Look for all greens: Maybe certain companies will come up as mostly green. It might be interesting to dive deeper into these

The body of the prompt is the framework we just explained.

The output is designed as a table, so one snapshot gives you an idea of how Buffett-like this company is.

You can copy/paste the prompt at the end of this article. You only need to change the company name. As always, make it your own, edit it.

You’ll see other articles like this appearing in the future. Let us know what you think:

Until soon, with more prompts and workflows!

Kevin

Share

The prompt:

# Your Role

You are an experienced equity analyst who has read everything Warren Buffett has ever written. For all intents and purposes, you are Buffett GPT. You search for the truth, weeding out the subjective from the objective. You use the maximum amount of information available to you in your research. I need your help in analyzing and understanding a company.

# The data

Fetch and integrate publicly available real-time data

Pull recent, credible, and relevant financial and strategic data about the company and its competitors from online sources (e.g. annual reports, earnings calls, news, financial platforms, patent databases, customer reviews, etc.).

# Your Task

The company we are researching today is called {Company_Name} which is publicly traded on the London Stock Exchange with the ticker Wise.

Can you go through the publicly available filings and break down the balance sheet, income statement, and cash flow statement as Buffett would do it?

Here’s some additional context to provide structure in analyzing these 3 statements.

For the balance sheet:

Buffett wants companies drowning in cash. So if cash and cash equivalents has been increasing over the last years, it’s a good sign the company is doing something right.

➡️Look for a steady rise in cash and cash equivalents

### Inventory

In the case of a company that holds inventory, a steady increase in inventory over the last years signals the company is selling more and more products. This is a good sign.

### Net receivables

The best companies have a cash conversion cycle that is negative. In other words they get paid in cash first and only have to pay their own suppliers later.

We dove deep into the cash conversion cycle in a previous article.

➡️Buffett looks for net receivables/sales that are lower than competitors '

### Long-term debt

Debt on itself is not an issue. Strong companies usually have a low debt level. But even if they have debt, look at the earnings power compared to the debt.

➡️Buffett looks for a long-term debt-to-earnings power ratio below 4 (if profits are cash, they can pay back debt in 4 years)

### Debt to equity

➡️Is debt to equity consistently low, preferably below 80%?

### Retained earnings

This is one number that gives you information about the history of the company. How much earnings have been retained within the company in the past?

➡️Buffett looks for companies with a history of increasing retained earnings

### Treasury shares

When a company buys back shares, it is registered as treasury shares. This reduced the amount of shares outstanding of the company.

➡️Buffett looks for companies with increasing treasury shares.

### Return on Equity

Has Return on Equity been stable or grown over the past decade? ROE is a measure of efficiency, although it might not be the best measure:

ROE can be a trap, it might be better to look at ROIC.

For the income statement:

Gross Margins

➡️The gross margin (gross profit/revenue) needs to be high, ideally above 40%.

It could signal pricing power. It provides a cushion for the company when things go sour.

Next, he takes a look at how different operating expenses compare to gross profits between different companies.

He’s looking for companies that are drowning in cash, so the cost structure matters.

SG&A

Selling, General & Admin costs can eat away at profits.

➡️He looks for SG&A/gross profits lower than 30%. The lower the better.

R&D

Phil Fisher loved R&D because it is needed to fuel future growth. But a company that can sustain its future growth with low R&D spending might have something special.

High R&D costs could mean the advantage is temporary, that it is not inevitable, or structural.

➡️Look for R&D/Gross profits < 10%

Depreciation

Depreciation is a cost. You already know he despises EBITDA.

Even worse, the cash was already paid in full at the start of the investment!

Low depreciation costs could mean that it's a capital-light business. And capital-light business might be more profitable.

➡️Look for Depreciation cost/gross profits < 10%

You don’t need the balance sheet. The Income statement can give you a signal of financial health.

High MOAT companies can have debt, but their earnings power is so strong that if they need to, they can pay it back pretty quickly.

➡️Look for Net interest/Operating profit < 10%

Net Income and EPS

Buffett wants a highly profitable company. He wants the company to have a history of rising profitability, not only on the amount but on the of earnings per share.

➡️Look for rising income and EPS

Income Before Tax and Normalization

There are no criteria for this, but I wanted to mention 2 important things:

- All investments are relative. Buffett likes the income before tax because if he divides it by the market cap he gets an equity yield. This allows for easy comparison to a bond yield or another company.

- When he talks about earnings, he means normalized earnings. We need to remove one-off profits or losses or even look at the cyclical nature of the business. Don’t just look at the P/E on a screen.

For the cash flow statement:

Profits are an opinion, cash is a fact

Alfred Rappaport

### CAPEX

Capital investment is needed to maintain and grow a business in the future. But here are big differences in industries.

➡️Because Buffett is looking for a capital-light business, he wants CAPEX/Net Income to be below 25%.

### Buybacks

We already covered this through the balance sheet, but the cash flow statement will also show the repurchase amount of shares (if any).

# Reasoning Instructions

“Base reasoning only on evidence provided or well-known industry facts; clearly state any assumptions.” Go through the different financial data 1 by 1.

# The Output Format

The output should be structured into 3 parts.

Buffett's balance sheet analysis

Buffett’s income statement analysis

Buffett’s cash flow statement analysis

Conclusion: Based on the above analysis, come up with a conclusion. Is this company high-quality based on the analysis of the financial statement?

Provide a checklist for all criteria. Present that checklist in a table. Use a traffic light (with actual colors presented by dots in green, yellow, and red) representation to visualize the value of each criterion. Clearly split the table into 3 parts: Balance Sheet, Income Statement, and Cash Flow Statement.

AI agents have been built and put into a room together.

And now they are talking to each other, building a church.

and constructing their own language so that humans can’t track what they are doing.

Sounds like sci-fi? Pretty crazy, right?

Besides it being cool, what we want is a 24/7 junior stock analyst who can help us make better investing decisions and save hours on research.

After spending a year experimenting with ChatGPT to build such an assistant, maybe this new, powerful software finally bring us closer to one.

This article is the first in a series where we’ll build our own analyst called Gordon. (Gekko)

As you can read from my telegram account, Gordon is eager to start working together.

In this article, you’ll learn:

1. What it is

2. How to set it up

3. Life-changing use cases

4. Stock analyst uses cases

5. Conclusion

Openclaw, what’s in a name

Initially launched as clawdbot (but Claude was not happy with that), changed into moltbot, and now called Openclaw, it’s an open source “AI-assistant’ project created by Peter Steinberger.

To simply show you what makes it different, here’s a comparison table with ChatGPT and other “assistants”:1

But let me give you a broad overview of how it works before delving into all of its features.

It works like this (in orange, you can see my current setup):

• Layer 1 (The channels): All the ways you can communicate with your assistant. You can configure multiple. I choose Telegram because it’s the simplest to set up, and I don’t use it for personal messaging (I use WhatsApp for that). You talk to your assistant and give them instructions

• Layer 2 (The gateway): This is the heart of OpenClaw. It manages all your message sessions, routes them across different channels, handles tool calls and skill execution, and most importantly, builds a persistent memory.

• Layer 3 (The LLM) of your choice or the brain. This is what makes it feel like a true assistant.

For the moment, Gordon has access to my Notion database and can search the internet through the API with the Brave browser.

Killer features

• Imagine you have a proactive assistant -> It will propose things to do, and if you agree, execute them while you sleep

• Remembers the past: Openclaw stores chats and previous messaging. It builds a living memory.

• Can actually do things (if you decode to connect it with your apps: email, calendar, Notion, etc.)

And people are building things, and some of the results have been wild. Just read this ⬇️

Ok, let’s backpeddle a bit.

How to get this thing to work?

How to set it up, and should you?

There are plenty of setup guides out there. I used this one, and it worked like a charm. (It took me about 2 hours the first time, as it was a long time ago, I used a terminal. I then restarted the entire process and got Openclaw up and running in 30 minutes. Note: I’m no developer. I consider myself a Power User. I feel comfortable with a command-line interface.

Be aware that this is not a “download a file and run an installer”. There’s a lot more that goes into this.

• You need to decide if you run it locally (Mac mini or other computer) or remotely with a virtual private server (I chose this)

• You need to be aware that this is still new. Security concerns have already surfaced. The guide I used does take that into account.

But even with these hurdles, it is compelling because it gives us a glimpse into the future and what a true AI assistant could look like.

And because it’s so new, there’s a lot of room to be creative. Figure things out.

Test, fail, and repeat.

But is it free?

No

I mean, a developer can run this locally with a local LLM like Llama. But most of us are not developers.

So if you’re like me, you’ll need to pay:

• For the VPS: I pay about 5 USD per month

• For the AI API: This depends on which one (Claude, OpenAI) you use, which model you use. I use GPT5 mini. I’ll let you know in the future how much this will cost me (although I have a GPT Plus subscription, API usage is separate from this)

All in all, if you want to experiment with it, this setup is cheap.

So, should you launch yourself and install Openclaw?

Here’s what to take into account:

• Security concerns have surfaced

• It’s not free, but pretty cheap

• You don’t need to be a developer, but be somewhat familiar with tech (don’t sweat, I use ChatGPT to help me. Worked like a charm.)

Whatever you decide, you can track my experiments with it right here at Invest with AI.

Subscribe now

Life-changing use cases?

Examples

1. Overnight coder

Build websites and track what OpenClaw does with a Kanban board

I see a lot of people use Claude as an LLM because it’s so good for coding. The above Kanban board (to track tasks) is not native to OpenClaw. It was created by prompting OpenClaw to track all the tasks OpenClaw is doing. This is a neat way to see which tasks you asked your assistant, and what the progress of those tasks is.

Now the kanban is peanuts. People are letting Openclaw vibe code tools while they sleep.

2. Reduce admin work

Manage your emails. Draft replies. Sort them. Manage your calendar. All those little time drainers, Claw can help you with this. There is specific software for this that already exists. You don’t need it anymore.

3. Competitive analysis

Scan your competitors. Receive a weekly summary of what is happening. Did they do anything incredible? Specific example: A YouTuber who tracks his 5 biggest competitors. Openclaw lists videos that have gone viral. What did they use as a thumbnail? Etc.

4. Automatic assistant

Let Openclaw take a look at your to-do list. Let it take tasks it can do automatically.

5. Personal CRM(Customer Relationships Manager)

Based on every email, text, or DM you have, create a database of your contacts and all your messages. Continuously update automatically. Take AI notes and create follow-up tasks

New skills

New skills are being added by the community. These are tools, API connections you can add to your assistant.

Here are some of them in the finance space:

Finance News Briefings

Market news briefings with AI summaries. Use when asked about stock news, market updates, portfolio performance, morning/evening briefings, financial headlines, or price alerts. Supports US/Europe/Japan markets, WhatsApp delivery, and English/German output.

Yahoo Finance

Get stock prices, quotes, fundamentals, earnings, options, dividends, and analyst ratings using Yahoo Finance. Uses yfinance library - no API key required.

Financial Market Analysis

Precision Financial Insights - Analyze stocks, companies, and market sentiment using authoritative data. Powered by Yahoo Finance and enhanced with intelligent news synthesis by we-crafted.com/agents/financial-market-analysis - Buy CRAFTED_API_KEY on our website to start using

Subscribe now

For stock analysis

So, I plan to use it to help me with stock research.

Someone used it to actually trade, but that didn’t go down so well. (could be a hoax though)

My goal in the coming weeks is the following:

Earning report analysis

Gordon’s first task:

• Scan for new earnings releases on the stocks in my portfolio and watchlist

• Fetch the new data

• Add it to the existing information on these stocks

• Run a specific prompt to analyze what happened and how the thesis might change or shift.

• Update the valuation model

• Propose a draft report I can edit

If Gordon can do this, it would save a couple of days every quarter.

Ideation engine

There are different ways to get ideas for investment opportunities:

• Stock screeners

• Your network or the people you follow

• Going A to Z through the stock market

That last one is a source for great ideas, but it takes time, a lot of time.

So the second task we’ll ask Gordon is to build an ideation engine A to Z

• Go through a stock exchange A to Z

• Use a specific pattern or framework that he needs to look for

• For each stock, write a summary and assessment + scoring

• Rank the investment candidates

And this brings us to a final idea.

Weekly brief: What are people buying?

Lots of people on the internet advise on what you should do with your money.

But what are they actually buying?

Just like tracking insider buys from management, we’re going to ask Gordon to:

• Scan newsletters and publications

• Flag when there is information that somebody has taken a sizable position in a stock (at least 10% of the portfolio)

• Provide a weekly summary report with a table

This brief will, of course, be heavily dependent on which data is publicly available. But I believe X can be a great source, as people often mention what they do (no guarantee that they actually did it, though)

Conclusion

These are exciting times.

With each day, we’re getting the tools to power up our investment analysis.

I will be diving headfirst into this and report back to you with practical use case examples.

If you want to follow how Gordon evolves as our assistant, I’ll keep you up to date!

Until soon!

Kevin

Share

The goal of this newsletter is simple: save hours on investing research.

How:

• By exploring the best prompts and workflows

• By reviewing the newest AI Tools and software

• By building our own “agent” to help us make better investing decisions

The “rise of AI” gives us 2 paths:

1. Embrace it, and stay on top

2. Ignore it and know that others are using it

If you’re reading this, I bet you’re in the first camp.

I’m glad you’re here.

Let’s take this journey together!

Consider this.

You look at 50 companies in a year. You pick 10 and build a portfolio.

What if you could study 500 companies in a year and pick 10 of them?

Would that last portfolio perform better?

It probably would.

Humans rank through comparison. You’re thinking of buying a new computer or car, and spend hours and hours comparing different models, to finally come to the best price-quality product you want.

Why not do the same thing with stocks? It looks like this:

But there’s a problem.

Your time is precious.

Maybe you have 1 hour, or maybe 8 hours every day for research. It doesn’t matter. You don’t want to spend hours on a stock, only to have to put it aside and say:

Nah, this is not for me.

Imagine your process is like mine: In 3 phases. You want to go fast through the first phase. Slower through the second. And once you finally find something interesting, only then do you go deep.

That’s efficiency.

The birth of LLMs now gives us a tool to accelerate our process. Especially the number of companies we can go through in that first phase.

Let’s fire up that research rocket!

This sequence of prompts will help you analyze any stock in the span of 15 minutes.

I picked Rocket Lab. It’s been on my list for a long time. Let’s dive in.

Subscribe now

Step 1: Breaking down the Business Model

The first step is to understand the business model and then decide whether it is in your circle of competence or too hard to understand. To break down the business model, we can use the following prompt and put it into ChatGPT:

# Your Role
You are an experienced equity analyst. You search for the truth, weeding out the subjective from the objective. 

# The data
Fetch and integrate publicly available real-time data. Pull recent, credible, and relevant financial and strategic data about the company and its competitors from online sources (e.g. annual reports, earnings calls, news, financial platforms, patent databases, customer reviews, etc.).

# Your Task
The company we are researching today is called Rocket Lab, which is publicly traded with the ticker RKLB.

I specifically want an answer to these 3 questions:

How does this company make money?
What does it need to do to generate its business?
Is it a capital-intensive business?

# Output Format
Please explain it to me like I’m a 10-year-old. 
Provide an example to illustrate the business model.

To see the complete prompts and feedback for all 5 steps, click here:

Complete Prompts + Feedback

ChatGPT provided us with the following output:

How does this company make money?

Rocket Lab makes money mainly in two ways:

a) Launch Services

Rocket Lab operates rockets (like Electron) that carry satellites into space for customers — both governments and private companies.

• Customers pay a fee for each launch (for example, around $7–8 million per Electron launch). (easymarketanalysis.com)

• These fees show up as revenue when a launch is completed.

b) Space Systems

Rocket Lab also designs, builds, and sells space hardware (like satellite platforms, components, and mission services).

• It gets paid to manufacture satellites and parts, often through multi-year contracts that give more predictable long-term revenue.

• Space Systems has grown to be the largest contributor to total revenue (about ~70% recently). (Forbes)

✔ Combined: These two segments bring cash into the company — launches pay when rockets fly, and space systems pay across manufacturing and service contracts.

What does it need to make money?

To generate revenue, Rocket Lab must do three key things:

a) Build and fly rockets

• Engineering and manufacturing of rockets like Electron and (in the future) Neutron.

• Launch infrastructure — launch pads, ground systems, range operations — must be maintained and operated. (Forbes)

b) Manufacture satellites and space hardware

• Production facilities, suppliers, and engineering teams must build satellite components, spacecraft platforms (e.g., Photon, Flatellite), and mission services. (easymarketanalysis.com)

c) Sign contracts and deliver on them

• Government and commercial agencies must commit to launching space systems contracts.

• Rocket Lab then executes those contracts — meeting timelines, performance expectations, and compliance for complex space missions.

Without engineers, factories, launch sites, and reliable execution, the company cannot deliver launches or space systems.

Is it a capital-intensive business?

Yes — very capital-intensive.

Rocket Lab spends huge amounts of money upfront on:

• Rocket R&D and manufacturing (materials, engines, structures)

• Launch pads and infrastructure (sites in New Zealand and the USA)

• Satellite production facilities

• Hiring engineers and specialists

These costs happen before revenue arrives from customers. That’s why Rocket Lab operates at a loss today — it spends more than it earns while growing its capabilities. (Value Sense Blog)

And let’s also show the “simple example” we asked it to give:

This is typical for space companies: high fixed costs and long development cycles. It’s analogous to needing a very expensive factory or a fleet of vehicles before you can start selling goods.

➡️The takeaway: So far, this business model is similar (but not quite) to aircraft manufacturers like Airbus and Boeing, but vertically integrated and smaller. The key difference would be that Rocket Lab also operates the vehicle, whereas Airbus mostly sells planes and lets airlines fly them. Although this isn’t a capital-light business (quite the opposite), it has high barriers to entry and only allows a few players in the market, which brings us to…

Step 2: Find out what the industry looks like

In this step, we look at what makes the company unique amongst its competitors and try to get a hold of what the industry looks like.

Here is the prompt:

Does their business model show signs of counter-positioning towards established businesses? How many other companies are there out there that are similar to Rocket Lab? If there are only a few, why is that? Are there any particular barriers to entry? Why does the establishment not replicate what this company is doing?

Note: This is a simple prompt, but you already primed the LLM with the first one, so this is good enough.

And here is the feedback

Counter-Positioning vs. Established Aerospace

a) Focus on Small Launch with Vertical Integration

Rocket Lab deliberately targeted a niche — small satellite orbital launches — which was underserved by the traditional aerospace players (like United Launch Alliance, Arianespace, or Russia/China state providers) that historically focused on:

• heavy or medium lifts,

• large government contracts,

• expensive, bespoke engineering projects.

b) Nimble Cost Structure & Tech-First Engineering

Unlike legacy players (e.g., Boeing/LMT via ULA, NASA contractors), Rocket Lab uses:

• modern manufacturing like 3D-printed engines,

• iterative software & automated design,

• private capital incentives,
  to reduce costs and increase launch cadence — not large, long multi-year procurement cycles. (Porter’s Five Forces)

How Many Real Competitors Exist?

• SpaceX — Dominates larger payload & rideshare launches globally. (Porter’s Five Forces)

• Blue Origin (New Glenn) — Aimed to take on heavier markets but has not reached mature cadence yet. (Le Monde.fr)

• United Launch Alliance (ULA) — Legacy provider with larger rockets, not small dedicated launches. (Wikipedia)

• Firefly Aerospace — Developing similar small-to-medium launch rockets. (Ars Technica)

• Relativity Space, Stoke Space, Isar Aerospace, Orbex, Astra, etc. — Early or development stage, many with no consistent commercial flights yet. (Wikipedia)

Why Are There So Few? Barriers to Entry

The commercial launch market is one of the most capital-intensive and technically difficult sectors in modern industry. The main barriers include:

• Massive Capital Requirements

• Technical Complexity

• Regulatory & Licensing Hurdles

• Economies of Scale & Data Advantage

• Market Demand Reality

Why Do Establishments Not Simply Replicate This Model?

a) Inertia & Legacy Business Models

Big defense/aerospace incumbents are structured around traditional programs and large contracts. They cannot pivot quickly or reduce costs without major organizational change — even if the idea works. (The Guardian)

b) Risk Tolerance & Culture

Rocket Lab and SpaceX operate more like Silicon-era tech engineering firms, tolerating iterative risk and quick learning. Legacy Aerospace has historically avoided failure at all costs. (Barron’s)

c) Scale & Strategic Focus

Large incumbents often rely on government budgets and strategic defense work, where price competition and innovation cycles differ vastly from the small-sat launch segment.

➡️The takeaway: Ok, so Rocket Lab has entered an industry with high barriers to entry, and it has signs of counterpositioning, with a niche focus within the market.

So far so good.

Step 3: Competitive Dynamics

To get more insights about how Rocket Lab will be able to defend its position and become one of the few players once the market matures, I will use Hamilton Helmer’s 7 Powers with the following prompt: (the actual prompt is huge). You can copy it through the button below.

Complete Prompts + Feedback

Evaluate the company against each of Hamilton’s 7 Powers:

1. Scale Economies
2. Network Economies
3. Counter Positioning
4. Switching Costs
5. Branding
6. Cornered Resource
7. Process Power

The feedback:

Scale Economies

• Assessment
  🟡 Moderate

• Competitive Context
  Compared to Astra or Isar Aerospace, Rocket Lab has superior scale and reliability. Compared to SpaceX, Rocket Lab is clearly subscale. Scale helps Rocket Lab survive — it does not yet dominate.

Network Economies

• Assessment
  🔴 Weak

• Competitive Context
  Unlike satellite constellations (e.g., Starlink) or software platforms, no launch provider in the industry enjoys true network effects. This is a structural industry limitation, not a Rocket Lab failure.

Counter Positioning

• Assessment
  🟢 Strong

• Competitive Context
  This mirrors SpaceX’s early counter-positioning vs. Boeing/Lockheed. Today, Rocket Lab occupies a similar role one tier down the payload stack. Most incumbents still haven’t meaningfully replicated this model.

Switching Costs

• Assessment
  🟡 Moderate

• Competitive Context
  Switching costs are higher than in commodity industries, lower than in enterprise software. Rocket Lab’s reliability record strengthens this power relative to newer entrants.

Branding

• Assessment
  🟡 Moderate

• Competitive Context
  Stronger brand than Astra, Firefly, and most European startups. Significantly weaker than SpaceX, which has become the default launcher in many customers’ minds.

Cornered Resource

• Assessment
  🟡 Moderate

• Competitive Context
  Compared to early startups, Rocket Lab has meaningfully cornered executional know-how. Compared to SpaceX, its resource depth is shallower and less defensible long-term.

Process Power

• Assessment
  🟢 Strong

• Competitive Context
  This is Rocket Lab’s strongest power, second only to SpaceX in the commercial launch sector. Legacy aerospace struggles here the most.

🔍 Bottom-Line Strategic Takeaway

Rocket Lab’s moat today rests primarily on counter-positioning and process power — not scale or networks.
That makes it defensible against new entrants, but still vulnerable to SpaceX’s overwhelming scale if market segments converge.

➡️The takeaway: Now that I have a good overview of what competition looks like and know that Rocket Lab’s biggest strength is the process power or execution, it is essential to screen the management in the next Step.

Step 4: Quality of management and its incentives

The prompt I use to assess the quality of management:

Decide if management’s incentives align with small shareholders. Render a verdict (Aligned / Mixed / Misaligned) with hard evidence. Score these 8 areas (0–2 each; 16 max):

Ownership & Skin-in-Game 
Pay Design 
Pay Metrics Fit 
Dilution & Equity Plan 
Capital Allocation Record 
Per-Share Outcomes 
Governance & Shareholder Rights 
Integrity & Culture 

The feedback:

🎯 Mixed — Management incentives only partially align with small shareholders.
There are meaningful stock incentives and insider ownership, but dilution risk, heavy equity compensation, and some structural governance features reduce alignment.

📊 Scorecard

TOTAL SCORE: 8/16 — Mixed alignment

➡️I cut some of the feedback to focus on the essential parts. In my investment book, the alignment seems good; the only thing bothering me is the SBC, but this should come down as a percentage of revenue in the future.

Subscribe now

Step 5: Understanding the investment thesis

The last step before deciding if I want to dive deeper or not is to understand what sentiment is driving the stock at the moment. This is essential as it will reveal what you have to believe or disbelieve in order to make the investment work.

The Prompt:

Create a one-page report on Rocket Lab (RKLB) that covers the following sections:

1) Recent Developments (3–5 bullets)

2) What Bulls Say (3–5 bullet points)

3) What Bears Say (3–5 bullet points)

The feedback:

Recent Developments

• Record Revenue & Margin Growth: For Q3 2025, Rocket Lab reported $155 M in revenue (+48% YoY) and a GAAP gross margin of 37%, its highest yet, with guidance of $170–180 M for Q4 2025. This reflects broad demand across launch services and space systems. (GlobeNewswire)

• Launch Cadence & Contracts: The company achieved 17 dedicated Electron launch contracts in Q3 and is on track to exceed 20 total launches in 2025, including HASTE hypersonic test missions. (GlobeNewswire)

• Strategic Acquisitions & Expansion: Rocket Lab completed the $325 M Geost acquisition, expanding capabilities into national security payloads, and continued M&A to strengthen its space systems portfolio. (GlobeNewswire)

• Neutron Rocket Milestones: Launch Complex 3 is operational for the medium-lift reusable Neutron rocket, with initial arrivals and qualification testing targeted for Q1 2026, moving toward a debut flight. (GlobeNewswire)

What Bulls Say

• Strong Revenue Growth & Backlog: Bullish investors highlight record revenues and robust backlog fueling growth visibility, with extended guidance and margin expansion as evidence of demand strength. (Q3 2025 $155 M, +48% YoY) (GlobeNewswire)

• Launch Dominance & Operational Reliability: Rocket Lab’s Electron rocket is a proven small-launch workhorse, with record cadence and industry-leading reliability, validating its execution excellence. (Reddit)

• Strategic Defense Contracts & M&A: Large government awards (e.g., U.S. Space Development Agency satellite contracts), plus the Geost acquisition, position Rocket Lab as a vertically integrated national security partner. (Investors)

• Medium-Lift Growth via Neutron: The Neutron rocket’s near-term debut offers a step change in addressable market and revenue per launch relative to small-sat missions. (Nasdaq)

What Bears Say

• Profitability Still Elusive: Despite revenue growth, Rocket Lab remains loss-making on Adjusted EBITDA and net income, with projected losses continuing in the near term. (Q4 2025 adjusted EBITDA loss expected) (GlobeNewswire)

• Capital Intensity & Dilution Risk: Heavy investment in Neutron, M&A, and space systems raises cash burn and share count dilution, potentially pressuring per-share metrics until profitability is firmly established. (GlobeNewswire)

• Competitive Pressure: The space launch market remains competitive, with incumbents like SpaceX dominating larger payload segments and exerting pricing pressure, limiting Rocket Lab’s pricing leverage. (Nasdaq)

• Neutron Execution Risk: The medium-lift Neutron rocket is a strategic pivot but carries developmental and execution risk, with timing and certification uncertainty potentially impacting valuations. (Nasdaq)

• Valuation Volatility: Despite strong rallies, the stock has exhibited significant volatility and may reflect speculative flows rather than fundamentals, with valuation multiples potentially stretched. (AInvest)

Conclusion

In 15 minutes, we’ve learned:

• What the business model looks like

• It entered a high barrier-to-entry industry with only a few but strong competitors

• It shows clear signs of counter-positioning towards the incumbents and process power

• Management incentives are “reasonable,” although, as is typical with these types of companies, we need to look at how SBC evolves

• Strong backlog and revenue growth, but profitability is still absent, while valuation remains stretched

I would complement such an analysis with some metrics from fiscal.ai.

It makes no sense to look at profits, as there are none. The only sensible metrics are EV/Sales or maybe EV/Gross Profits complemented with the expected growth.

• EV/Sales of 84 divided by 42% expected growth = EVSG = 2

That’s expensive. (Note for ASTS, the same calculation gives us 2.7 - even more expensive)

Rocket Lab is an interesting company. It is within my circle of competence. As an engineer, I would love to own this company, but it’s too expensive.

We’ve also seen that one powerful feature is to use AI early in your research funnel.

This first article was the appetizer. We’ll apply prompts from physics (Richard Feynman) and psychology (mental models) to go faster and deeper inside these companies.

We’ll look at the newest tools, like Clawdbot (or is it Moltbot 😉), and check if they can help us save time and make money.

Until soon!

Kevin

Share