Cookbook

Cookbook — Anthropic's reference recipes

You are about to build something Anthropic already published a reference implementation for. Read theirs first.

This skill routes into 94 recipes from anthropics/claude-cookbooks (MIT), pinned to a recorded commit in tools/cookbooks/COOKBOOKS.lock. The index ships with OmegaOS, so search always works. The notebooks themselves are an optional local corpus.

The rule this exists to enforce

R-SKILL-ATLAS says: find the real thing and run it, never answer generically. This skill is that rule applied to Claude-API engineering. When a mission needs a RAG pipeline, an eval harness, a citations layer, an orchestrator-workers fan-out, contextual embeddings, prompt caching, a moderation filter or a cost-optimization pass, there is a reference implementation and re-deriving it from memory is the failure mode, not the work.

1. Find the recipe

omega-skills --rag "<the need, in plain language>"

Cookbook rows are labelled by where they live, and the label is honest:

For the machine-readable form (carries local and url):

omega-skills --rag "<need>" --json

Browse everything by category in the Skill Atlas — the Anthropic cookbooks tab of ~/.omega/artifacts/omega-skill-atlas.html.

2. Read it

Corpus present (~/.omega/cookbooks/): read the notebook directly. It is JSON — pull the source cells rather than paging the whole file:

python3 -c "
import json,sys
nb=json.load(open(sys.argv[1]))
for i,c in enumerate(nb['cells']):
    src=''.join(c['source'])
    if src.strip(): print(f'--- [{i}] {c[\"cell_type\"]} ---\n{src}\n')
" ~/.omega/cookbooks/<path>.ipynb

Corpus absent: WebFetch the url from the --json output. It is pinned to the recorded commit, so it matches the indexed description exactly.

Install the corpus for offline work:

tools/cookbooks/install-cookbooks.sh          # ~99M, ~3s, images excluded
tools/cookbooks/install-cookbooks.sh --status

3. Adapt it — do not paste it

The recipes are Python against the Claude API with an ANTHROPIC_API_KEY. OmegaOS is Rust-first and runs on the Claude Code subscription (R-STACK). So:

• Take the method — the chunking strategy, the judge rubric, the retry shape, the cache-breakpoint placement, the eval design. That is the value.
• Port it to the OmegaOS stack. A recipe is a reference, never a dependency.
• If you genuinely need the Python path, it is a documented exception (R-STACK) and the key lives in ~/.omega/secrets/, never the repo (R-ENV).
• Model IDs in the notebooks age. Check them against the claude-api skill before running anything.

4. Cite what you used

R-CITE: name the recipe and its path in your report, so the next agent can follow the same trail instead of re-deriving it.

Ported the contextual-retrieval chunking from capabilities/contextual-embeddings/guide.ipynb (cookbook @ 35f2eec).

Where each need lands

Keeping the pin current

tools/cookbooks/install-cookbooks.sh --update    # move the pin to upstream main
tools/cookbooks/install-cookbooks.sh             # fetch it
tools/cookbooks/build-index.py ~/.omega/cookbooks -o tools/cookbooks/recipes.json
python3 scripts/omega-skills-atlas.py && python3 scripts/omega-skills-rag.py build

Commit COOKBOOKS.lock and recipes.json together — the pin and the index are one artifact, and a drift between them means the RAG describes a recipe that the local notebook no longer matches (L0).

Also vendored from this upstream

Three recipes were substantial enough to become real OmegaOS skills rather than references. They are installed and reachable by name:

• analyzing-financial-statements — ratio analysis over real statements
• creating-financial-models — DCF and sensitivity analysis
• applying-brand-guidelines — programmatic brand compliance on generated docs

And docs/AGENT-PATTERNS.md carries the five canonical agent workflow patterns plus Anthropic's production research-lead / subagent / citations prompts, mapped onto R-GRAPH and R-ORCH.