The Anthropic Cookbook is a GitHub repository of Jupyter notebooks, each demonstrating a technique with Claude. It sits alongside the OpenAI Cookbook as the reference collection for its ecosystem, and the comparison is fair because they serve the same purpose in the same format with similar strengths. What makes cookbooks valuable is that everything runs. Documentation describes a capability.
A cookbook notebook implements it, calls the API, and shows you the output. That difference matters more than it sounds, because the gap between understanding a feature in the abstract and getting it working is where most people lose time. Opening a notebook, running it, seeing a result, then modifying it for your own case is an efficient path from reading to building. The content selection reflects what people actually get stuck on.
Tool use gets proper treatment, including how to define tools clearly, how to handle multiple calls, and how to deal with a model that picks the wrong tool. Retrieval augmented generation is covered with working implementations rather than architectural diagrams. Structured output, where you need parseable JSON reliably rather than usually, is handled with the specific techniques that make it work. There is material on sub-agent patterns, on vision, on long context strategies, and on evaluation.
The evaluation notebooks deserve particular attention because evaluation is the most commonly skipped step in LLM development and having a working example lowers the barrier considerably. The provenance is worth something. These are written by people at the company that trained the model, who know how it behaves and where it is weak. When a notebook says a particular prompt structure works better, that recommendation comes from more evidence than a blog post from someone who tried it twice.
There is a natural caveat, which is that a vendor's cookbook shows the vendor's model working well, and the technique explanations have generally been honest about limitations in my reading. The patterns generalise. Prompt structuring, tool definition, retrieval design, evaluation approach. The specific syntax is Claude's and the thinking transfers to any model with comparable capabilities.
Someone who works through the tool use material will be better at tool use everywhere. What this is not is a course. There is no order, no prerequisite structure, no progression, no explanation of fundamentals. You need to already know what you are looking for.
Someone new to language model development should read the Anthropic documentation properly, or take a structured course, and then come here when they have a specific problem. Arriving with no orientation produces browsing rather than learning. The Python assumption is real. Notebooks, environment setup, API keys, dependency management, and enough Python to modify an example meaningfully.
Not advanced, and not nothing. Running the notebooks costs money. Not much for most of them, and enough to notice if you work through a lot of the retrieval and evaluation material, which makes many calls. Set a spending limit before you start rather than after.
Coverage is uneven, which is the nature of a collection that grows by contribution. Some areas have several thorough notebooks and others have one thin example or none. Checking what exists before planning to learn something from here saves disappointment. My four point one is for genuinely practical, runnable, free material from an authoritative source, covering the techniques people actually need, marked down for having no curriculum and for assuming a starting point that a beginner will not have.
If you build with Claude, this repository should be bookmarked. If you build with any model, the tool use and evaluation notebooks are worth your time regardless.