AI Coding Reference
Cookbooks, prompt registries, skill directories and catalogues. Everything here is material you consult rather than software you run: no account, no pricing tier, nothing to install.
How to Use This Hub
Start from the model you are already calling. If you write against Claude, the Cookbooks section is the fastest route to a working request. If you are looking for something to install into an agent, go to Skill Directories, which is where the catalogues live. The Specs and Protocols section is the background reading the rest of it assumes.
Cookbooks and Recipe Repositories
Vendor-published notebooks and worked examples. You read them and copy from them; there is nothing to install and no account to open.
Claude Cookbook Review
Anthropic's notebooks for tool use, MCP integration, agent patterns, extended thinking, and prompt caching.
Read guideGemini Cookbook Review
Google's quickstarts and notebooks for multimodal prompting, function calling, and grounding with Search.
Read guideOpenAI Cookbook Review
The oldest of the three, and still the widest: embeddings, evals, structured output, and retrieval recipes.
Read guideSkill Directories and Marketplaces
Places you go to find agent skills. Each one indexes other people's work, which is why they sit here rather than in the skills directory.
Skills.sh Review
An open directory and leaderboard for agent skills, with one-command installs across Claude Code and Cursor.
Read guideOpenAI Codex Skills Catalog Review
OpenAI's own repo of skill folders for Codex, covering deployment, design implementation, and testing.
Read guideClaude Skills Guide
What a SKILL.md file actually contains, how Claude loads one, and when writing your own beats installing someone else's.
Read guidePrompt and Chain Registries
Versioned stores of prompts, chains, and agent configurations. Useful as a starting point, rarely useful unedited.
Specs and Protocols Worth Bookmarking
The open formats and protocols the rest of this directory is built on. Read once, refer back often.
AGENTS.md Guide
The file-format convention for telling any coding agent how your repo works, and what belongs in it.
Read guideWhat Is MCP?
The Model Context Protocol explained: what a server exposes, what a client consumes, and why it caught on.
Read guideBest MCP Servers for Coding
The four worth installing first, and what each one actually adds to a session.
Read guideTracked References
Cookbooks, registries and catalogues we have reviewed and indexed. None of these is a product: there is nothing to buy, and several of them are directories of other people’s work. Each card links to the full write-up.
Claude Cookbook
referenceOfficial collection of Jupyter notebooks and code recipes from Anthropic for building with the Claude API. Covers tool use, MCP integration, agent patterns, extended thinking, RAG, prompt caching, multimodal capabilities, and production patterns.
Source: Cookbook Platform
Read the write-upGemini Cookbook
referenceOfficial collection of guides, quickstarts, and Jupyter notebooks from Google for building with the Gemini API. Covers multimodal prompting, function calling, grounding with Google Search, context caching, audio/video understanding, and agent patterns.
Source: GitHub
Read the write-upLangChain Hub
referenceCommunity prompt and chain repository integrated with LangSmith. LangChain Hub lets developers discover, share, and version-control prompt templates, agent configurations, and chain definitions: a prompt registry that connects directly to the LangChain ecosystem for rapid AI prototyping.
Source: Hub
Read the write-upMCP Market Skills
referenceAgent skills directory with focus on Claude, ChatGPT, and Codex. Discover, retrieve, and install reusable AI agent capabilities with step-by-step installation guides for multiple AI platforms.
Source: Website
Read the write-up
OpenAI Codex Skills Catalog
referenceOfficial open-source catalog of Agent Skills for OpenAI Codex. Skills are reusable folders of instructions, scripts, and resources that extend Codex with task-specific capabilities like deployment, design implementation, security analysis, and testing.
Source: Website
Read the write-up
Skills Directory
referenceOpen registry for verified agent skills with 3,500+ skills. Discover and install quality-focused skills for Claude Code, Cursor, Windsurf, and other AI coding agents with community verification.
Source: Website
Read the write-upSkills.sh
referenceOpen directory and leaderboard for reusable AI agent skills. Discover and install modular capabilities that enhance AI coding agents with procedural knowledge through simple one-command installation.
Source: Website
Read the write-upSkillsMP
referenceMassive agent skills marketplace with 71,000+ skills compatible with Claude Code, OpenAI Codex CLI, and ChatGPT. Uses the open standard SKILL.md format with smart search and category filtering.
Source: Website
Read the write-upReading the Reference Layer in 2026
Why a Reference Directory Exists at All
A tool directory answers one question: which of these should I pay for and install. That question is meaningless for a cookbook. Nobody chooses between the Claude Cookbook and the Gemini Cookbook the way they choose between Cursor and Zed; you open whichever matches the API you are already calling, and often both. Grouping them separately means the ranking signals that make sense for products (pricing, free tier, mobile output) stop being applied to things that have none of them, and the questions that do matter for reference material (how current is it, who maintains it, how much of it is runnable) get room to be asked.
The Catalogue Layer Nobody Planned
In 2025 an agent skill was a file you wrote yourself. By 2026 there are directories indexing tens of thousands of them, leaderboards ranking them, and marketplaces selling access to them. That layer appeared fast and it appeared without anyone designing it, which is why the vocabulary is still a mess: the same word covers the file, the folder, the registry, and the site that indexes the registry. Keeping the catalogues in one place and the skills in another is the smallest fix that stops the confusion spreading.
How to Read a Cookbook Without Copying Its Bugs
Vendor cookbooks are written to demonstrate an API, not to survive production. The notebooks skip error handling, hardcode model names that get deprecated, and assume a single user with a warm cache. That is fine for what they are, and it becomes a problem the moment a recipe gets pasted into a service. Read them for the shape of the call and the order of the arguments. Take the retry logic, the timeout policy, and the cost ceiling from somewhere else.
When a Registry Is Worse Than Writing It Yourself
A prompt registry pays off when the prompt encodes something you would not have worked out alone: a chain of tool calls, a scoring rubric, an evaluation harness. It costs you when the prompt encodes someone else's codebase conventions, because you inherit assumptions you cannot see and debug them later as mysterious quality drops. The rule of thumb is that generic technique travels and repo-specific instruction does not. A registry entry that mentions a directory layout is telling you it was written for a project that is not yours.
Frequently Asked Questions
What counts as a reference entry?
Anything you consult rather than run. Cookbooks, registries, catalogues, and directories all share the same shape: there is no account, no pricing tier, and no service running anywhere. You open it, take what you need, and close it. If there is something to install or sign up for, it belongs in the tools directory instead.
Why are these not in the tools directory?
Because listing them alongside Cursor and Lovable produced comparisons that made no sense. A directory of skills has no pricing to compare, no editor to benchmark, and no free tier to weigh up. Those pages existed because the entry carried the wrong label, not because anyone wanted to read them.
Why are skill directories here and not under Skills?
Because a directory of skills is not a skill. Skills.sh, SkillsMP, Skills Directory, and MCP Market Skills all index other people's work: you install what they list, not the thing itself. Filing them under Skills would put a catalogue in the same grid as the items it catalogues, which reads as sloppy the moment anyone looks twice.
How do I tell a catalogue from a product?
Ask what you would be paying for. A product sells you a runtime, an editor, a hosted service, or a seat. A catalogue sells you nothing: its value is that someone else did the indexing. LangChain Hub is the awkward middle case, because it lives inside LangSmith, which is a product. The Hub itself is still a registry you browse.
Do any of these cost money?
The cookbooks and open directories listed here are free to read. Some of them point at paid APIs: a Claude Cookbook notebook still needs Claude API credits to run, and LangChain Hub sits behind a LangSmith account. The reference itself is free; what it teaches you to call may not be.
How often is this list updated?
Entries are reviewed on the same cycle as the tool directory. Reference material moves slower than products do: a cookbook gains notebooks, a directory gains listings, but the thing itself rarely changes shape. Each detail page carries its own last-reviewed date.
What happened to the comparison pages for these entries?
They stopped generating and now redirect to the entry they were about. A page comparing a cookbook head to head with an IDE was never a page anyone needed, so removing it is the correction rather than a loss. Every write-up still lives at its original address.