Claude Cookbook: Anthropic's Official Playbook for Building with Claude

Last updated: October 12, 2026 - re-checked against the live cookbook index (97 entries now) and added the newest recipes.

Anthropic has been quietly building a comprehensive recipe collection. The Claude Cookbook is the company's official library of practical guides and examples, with 90+ entries spanning everything from basic vision setup to async multi-agent orchestration, programmatic tool calling, and Fable 5 fallback billing. It landed on the HN front page with 223 points and 117 comments - and the discussion was as much about the value of such resources as the cookbook itself.

What the Claude Cookbook Contains#

The cookbook lives at platform.claude.com/cookbook/ and is organized by category: tool use, agent patterns, evals, multimodal, RAG, skills, integrations, and the Claude Agent SDK. Each recipe has a clear title, author credit, category tags, and publication date.

Some of the most notable recent additions include:

• Programmatic Tool Calling (Nov 2025) - reduces latency and token consumption by letting Claude write code that calls tools within a code execution environment, rather than calling them via the standard tool-use loop
• Agentic Search Benchmark Reproduction (Jun 2026) - Mengting Li's walkthrough for building a Messages API harness that reproduces published DeepSearchQA and BrowseComp scores using programmatic tool calling, server-side compaction, and task budgets
• Async Multi-Agent Orchestration (Jun 2026) - Paul Chen's patterns for fixed N-agent teams with peer messaging and dynamically spawned async subagents
• Fable 5 Fallback Billing Guide (Jun 2026) - detecting safety classifier blocks and falling back to Opus 4.8, including streaming behavior and the pricing changes
• Context Engineering Tools (Mar 2026) - Isabella He's comparison of memory, compaction, and tool-clearing strategies for long-running agents
• Multi-agent teams under latency pressure and budgets (Sep 2026) - the newest entry in the Agent Patterns category as of this check
• Build a scheduled repository reviewer (Aug 2026) - a Claude Agent SDK recipe for recurring code review
• Claude Skills Collection (Oct 2025) - Alex Notov's series on building custom skills for Excel, PowerPoint, PDF workflows, and financial dashboards

What is striking is the breadth. The same page hosts Alex Albert's foundational guides from 2024 (extended thinking, tool choice, JSON mode) alongside cutting-edge June 2026 entries from engineers like Paul Chen and Mengting Li. It is a living document that has grown with the platform.

The cookbook also accepts community contributions through GitHub.

What HN Is Saying#

The HN discussion, at 117 comments and counting, split into several distinct threads. The most-upvoted comment (by mindwok, top of the thread) questioned whether any "how to use AI" resource is useful: "I'm either going to ask the AI how to do it, or if it's about using the AI then we can just bake it into the harness or wait for Anthropic/OpenAI to do it for me because they're always trivial."

This sparked a deep debate about prompt engineering. saberience argued that "the models are at this point smarter than you are, so the idea that you can prompt them 'better' is laughable really when discussing frontier models." Several commenters pushed back. Yiin shared a concrete counterexample: "Fable prompted to do a review found surface level issues, while prompting along the lines of 'assume it's wrong, prove it's correct' found much more in depth and real issues."

Another thread critiqued a specific recipe - Prompting for Frontend Aesthetics. semiquaver noted "the before and after images ... are hilarious. Did no one look at these to make sure the skill actually improved the design?" Sverigevader agreed: "Personally I prefer the before shots all the way down." This critique resonated - the HN crowd holds a low tolerance for marketing-optimized examples.

A more practical thread discussed CLAUDE.MD files. mexicocitinluez argued for minimal agent configs: "I think the best CLAUDE.MD is no CLAUDE.MD at all." rowanseymour noted the shift: "It used to be standard practice to let Claude scan everything once and describe your repo ... But now I think the tooling has gotten so good at just grepping around your repos, and maintaining memory from previous sessions, that the best practice is only use CLAUDE.MD for things that aren't obvious from the code."

Several commenters also joked about expecting actual food recipes from the title - a recurring pattern whenever "cookbook" enters tech naming. simonw shared that he has been cooking with LLMs "a few times a month for over a year now" and that "it's worked out well 9/10 times."

Why It Matters#

The Claude Cookbook serves a specific purpose that raw model intelligence does not replace. Knowing that programmatic tool calling exists, understanding when to use context compaction vs. memory, or recognizing that you can build an async multi-agent system with a shared hub - these are design patterns, not prompts. The cookbook catalogs what the Claude platform can actually do, which evolves faster than any single developer can track.

The counterargument from the HN thread - that these patterns will be absorbed into the model and harness over time - is partially correct. Anthropic does bake common workflows into the product. But the cookbook's role shifts as this happens. Older recipes become historical reference. New ones cover the frontier before it becomes productized. The Fable 5 fallback billing guide from June 2026 is a good example: it documents a real pain point that the platform had not fully automated at launch.

The existence of the cookbook also signals something about Anthropic's developer relations strategy. Unlike the fragmented landscape of third-party tutorials and community examples, the cookbook gives Anthropic a direct channel for opinionated, maintained guidance. It competes for developer mindshare with the OpenAI Cookbook (which predates it by about 14 months, as simonw noted) and the growing ecosystem of community resources.

For developers building on Claude, the cookbook is worth a browse at least quarterly. The agents you built last quarter may have simpler or cheaper implementations now. And the "Prompting for Frontend Aesthetics" recipe - critique notwithstanding - captures real constraints around Claude's stylistic defaults that anyone shipping AI-generated UIs should understand.

Sources#

• Claude Cookbook - Anthropic's official guides and examples. Accessed 2026-07-24.
• HN Discussion: Claude Cookbook - 117 comments, 223 points. Accessed 2026-07-24.
• Claude Cookbook GitHub Repository - Community contribution guide.
• OpenAI Cookbook - referenced in HN thread by beklein.

Continue Reading#

• Claude Code Skills Marketplace Launch - Anthropic's approach to modular, shareable agent capabilities
• Context Engineering Guide - managing context limits in long-running workflows (covered in the cookbook's context compaction recipe)
• Claude Agent SDK vs LangGraph - understanding the orchestration layer the cookbook builds on
• Prompt Engineering for Coding - the evolution of prompting strategies since 2024
• Extended Thinking in Production - one of the cookbook's foundational topics, covered by Alex Albert
• Claude Opus 4.5: Anthropic's Most Intelligent Model
• Claude - models, pricing, and API surfaces on one page
• Intro to Agents 101 course - from a first tool-calling loop to a production agent, the next step after cookbook recipes