claude-cookbooks is Anthropic’s pile of Jupyter notebooks for building with the Claude API. 40k stars, MIT, mostly Python. Not a tutorial series. Reference code you run, diff against your own, and copy from.

Worth pulling up: prompt caching, tool use with SQL and a calculator, sub-agents (Haiku feeding Opus), PDF parsing, JSON mode, moderation filters, and the eval notebooks. The evals are the quiet one. Most of us ship prompt changes on vibes. These show how to score a change against a held-out set before it merges.

The RAG notebooks are worth reading even if you’ve built RAG before. Pinecone plus Voyage embeddings, without hand-waving over chunking or retrieval quality. Multimodal is similar: vision, chart reading, form extraction. All executable, prompts sitting right there in the cell where you can mess with them.

Sub-agents is the pattern I’ve actually stolen. Haiku does the grunt work, Opus does the judgment, and the notebook is concrete enough that porting it into your own orchestrator isn’t a guessing game. Cheaper than running Opus end-to-end. Quality drop is small if you split the work right.

If you’re on the Claude API and haven’t skimmed this, you’re probably rebuilding something that already exists in a notebook with tests attached.