extract-recipe

Extract the sequence of prompts ("recipes") used to build artifacts with Claude Code.

Reads prompt history from ~/.claude/history.jsonl and resolves pasted content references from ~/.claude/paste-cache/.

Developed with Claude Code 2.1.45 using mostly Claude Opus 4.6, for Claude Code 2.x format history.

The prompts used to develop this tool are in recipe.md. Run make recipe to regenerate it.

Installation

Requires Python 3.9+. No external dependencies.

cd recipe-extraction

# Create and activate a virtual environment
/usr/bin/python3 -m venv .venv
source .venv/bin/activate

# Upgrade pip (required — the system pip 21.x cannot do editable installs
# from pyproject.toml-only projects)
pip install --upgrade pip

# Install in editable mode
pip install -e .

Usage

# List all projects with prompt and session counts
extract-recipe --list

# Save a recipe for sharing (redact sensitive paths/keys, write to file)
extract-recipe -r -t "My Project" . -o recipe.md

# Extract recipe by substring (matches if unique)
extract-recipe foo

# Use path components to disambiguate similar names
extract-recipe foo/bar      # substring: won't match foo2/bar
extract-recipe -e bar       # exact final component: won't match bar2

# Full path always works
extract-recipe /path/to/project

# Extract all projects at once
extract-recipe -a

# Redact home paths and API keys before sharing
extract-recipe -r myproject

# Keep system-generated boilerplate (stripped by default)
extract-recipe --raw myproject

# Extract as JSON to a file
extract-recipe --format json -o recipe.json myproject

# Browse with rendered markdown using glow
extract-recipe myproject | glow -p

# Check for potential names/identifiers before sharing
extract-recipe --audit myproject

# Copy default config for editing
extract-recipe --init-config

# Use custom pattern config
extract-recipe --config my-patterns.conf -r myproject

# Use a custom Claude config directory
extract-recipe --claude-dir /other/path --list

Project Matching

The project argument is matched flexibly:

1. Full path — exact match, always works
2. Substring — case-insensitive substring match against the full project path. myproject matches .../projects/myproject. Including / in the argument naturally narrows results by requiring component boundaries — e.g. foo/bar won't match .../foo2/bar. If multiple projects match, all are listed so you can refine.
3. Exact component (-e) — matches only projects whose final path component(s) are exactly the argument: -e bar matches .../bar but not .../bar2
4. Fuzzy suggestions — if nothing matches, similar project names are suggested (handles typos)

Boilerplate Stripping

By default, system-generated lines injected by AI coding tools are stripped from prompts. These are not user-authored content — for example, Claude Code appends transcript file references to plan-mode prompts in history.jsonl.

Detection uses regex patterns defined in boilerplate.conf (shipped with the package). The configuration has four sections:

• [strip] — regexes removed from prompt text (e.g. .claude/projects/.../*.jsonl paths)
• [skip] — prompts matching these are omitted entirely (e.g. /config, /login)
• [plan] — prompts matching these are collapsed to a title summary
• [redact] — pattern = replacement pairs applied when -r is used (e.g. home paths → ~, API keys → [REDACTED])
• [audit-stopwords] — case-insensitive regexes for common words to filter from --audit output (e.g. add(ed|ing)?)

Processing order: skip → strip → plan → redact.

If a user config exists at ~/.config/extract-recipe/patterns.conf, it replaces the package defaults entirely. Use --init-config to copy the defaults there for editing. Use --config FILE to load from a different path. Patterns use Python re syntax.

Use --raw to preserve the verbatim recorded text. --raw and --redact are independent: --raw --redact keeps boilerplate but redacts sensitive content within it.

Redaction

--redact applies pattern-based substitutions for common categories of sensitive content: home directory paths, API keys (AWS, GitHub, Anthropic, OpenAI, Google), UUIDs, and /tmp paths. Timestamps are replaced with sequential numbering (Session 1, Prompt 1.1, etc.).

Redaction is not exhaustive. It catches known patterns but cannot detect all sensitive content — names, project details, URLs, and other identifying information in your prompts are not automatically redacted. Use --audit to review your prompts for proper nouns and other potentially sensitive words before sharing, and consider manual review or additional tools for thorough anonymisation.

CLI Reference

extract-recipe [--claude-dir DIR] [--format {markdown,json}] [--list] [--audit] [-a] [-e] [-r] [-R] [-t TITLE] [--config FILE] [--init-config] [-o FILE] [project]

Flag	Description
project	Project path, substring, or path suffix to match
--list	List all projects with prompt/session counts
-a, --all	Extract recipes for all projects
-e, --exact	Match by exact final path component(s) instead of substring
--audit	List potential proper nouns in prompts (for manual review before sharing)
-r, --redact	Redact known sensitive patterns (home paths, API keys); not exhaustive
-t, --title TITLE	Override the project name in the output header (default: project path)
-R, --raw	Preserve raw prompt text (don't strip system-generated boilerplate)
--config FILE	Pattern config file (default: ~/.config/extract-recipe/patterns.conf)
--init-config	Copy default patterns to user config location for editing
--format	Output format: markdown (default) or json
-o FILE	Write output to file instead of stdout
--claude-dir	Claude config directory (default: ~/.claude)