Claude Cookbook

TL;DR Highlight

Anthropic published a cookbook of recipes for agents, tool calling, RAG, and caching with Claude

Who Should Read

Backend developers building agents or RAG on the Claude API who want official example code to reference

Core Mechanics

Claude Cookbook is a collection of examples
- It gathers practical guides and examples for using Claude, grouped by category.
- Categories include Agent Patterns, Tools, RAG & Retrieval, Evals, Observability, Multimodal, Skills, Thinking, and Fine-Tuning.
Recipes for tool-calling efficiency
- Programmatic tool calling (PTC) has Claude write code that calls tools in a code execution environment, cutting latency and token use.
- Tool search with embeddings uses semantic embeddings to discover tools dynamically, scaling to thousands of tools.
Recipes for long-context management
- Automatic context compaction compresses conversation history for long-running agents.
- Session memory compaction compacts instantly using background threading and prompt caching.
Recipes for cost and speed
- The Message Batches API recipe claims a 50% cost reduction via asynchronous processing.
- Speculative prompt caching warms the cache while users type to reduce time-to-first-token.

Community Reactions

Opinion that official guides are pointless
- Commenters said most of this is trivial, will be baked into harnesses, and agent workflow material feels like theatre.
- Another view was that such recipes get absorbed into the next model version.
Criticism of the frontend aesthetics examples
- Commenters said the before and after images are both bland, or just bland with gradients.
- Others pointed out even basic UI like table spacing is off.
Maintenance worries and alternatives
- One commenter said Copilot and OpenAI cookbooks started with many examples and later went unmaintained.
- The OpenAI Cookbook was suggested as another resource.
Opinion that the best CLAUDE.md is none
- One commenter argued tech stack and test style should be obvious from the code, so CLAUDE.md isn't needed.

How to Apply

Find one pattern you use in the Cookbook
- If your agent has many tools or high token use, read the PTC and Tool search with embeddings recipes today.
- Note what differs from your current implementation.
Add context compaction to long agent sessions
- Try the Automatic context compaction recipe on an agent with long conversations.
- Compare token counts and answer quality before and after.
Apply caching and batching to repeated calls
- Move non-realtime bulk requests to the Message Batches API.
- For latency-sensitive paths, measure time-to-first-token with the Speculative prompt caching recipe.

Try It

copy & run

# Paste into Claude (simple prompt for choosing a Cookbook recipe)

I run a service on the Claude API.
{{service description}}

My current problem is:
{{latency, cost, context overflow, too many tools, etc.}}

Prioritize which of these Claude Cookbook recipes to apply first.
- Programmatic tool calling
- Tool search with embeddings
- Automatic context compaction
- Speculative prompt caching
- Batch processing with Message Batches API
- Contextual retrieval

For each recipe, give one line on why and a 30-minute experiment.

Next: open the notebook for the top recipe and compare it against your code.

Terminology

Programmatic tool callingInstead of calling tools one by one, Claude writes code that calls them and runs it in an execution environment. It's like handing over a script rather than giving errands one at a time.

Context compactionSummarizing long conversation history to stay within the context limit. Like replacing meeting transcripts with minutes.

Contextual retrievalA RAG technique that prepends surrounding context to each chunk before embedding. Like labeling each book page with its chapter title.

Managed AgentsAnthropic's hosted agent runtime that manages sessions, environments, and memory.