Keyword-triggered context injection for AI coding agents. inspired by SillyTavern's world-info

CLAUDE.md files grow large, but most instructions are irrelevant to any given prompt.

Lorebook injects only matching entries via XML tags before user prompt.

It does not break the prompt cache!

curl -fsSL https://raw.githubusercontent.com/rolandmarg/lorebook/main/install.sh | bash

Downloads the binary, wires a Claude Code UserPromptSubmit hook, creates ~/.claude/lorebook/ with a starter entry.

Lorebook works with OpenAI Codex CLI too — same hook protocol. Add to ~/.codex/hooks.json:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "lorebook match"
          }
        ]
      }
    ]
  }
}

Both agents share entries from .claude/lorebook/ (project) and ~/.claude/lorebook/ (global).

Markdown files in .claude/lorebook/ (project) or ~/.claude/lorebook/ (global):

# ~/.claude/lorebook/lorebook.md
---
keys: [lorebook]  # trigger keywords
---
Lorebook injects context into prompts when
keywords match. Create .md entries with YAML
frontmatter: keys, exclude_keys, priority,
enabled. Case-insensitive, word-boundary
matching. `lorebook test` to verify.
https://github.com/rolandmarg/lorebook

Entries can inject content from existing project files instead of embedding it inline:

# .claude/lorebook/design.md
---
keys: [design, philosophy]
inject_files: [PHILOSOPHY.md, IDEAS.md]
---
Optional preamble text.

File paths are relative to the project root. Content from referenced files is appended to the entry body. Missing files are skipped with a stderr warning.

lorebook test "your prompt" # shows what would match
lorebook list               # shows all entries

MIT