GitHub is the first place anyone looks for example Claude skills. Searching "claude skills github" is the standard reflex after reading about the format. It's where most people then lose an hour.
A skill, in the format Anthropic shipped and documented at code.claude.com/docs/en/skills, is a folder. Inside lives a SKILL.md with YAML frontmatter and markdown instructions, plus any supporting scripts or templates. That's it. The format is intentionally simple, which is great for authors and brutal for discovery, because the bar to publish is "create a folder, push to GitHub." No app store. No review. No built-in rating.
So the practical question: where on GitHub is the good stuff, and how do you tell it from the rest before you git clone something into ~/.claude/skills/ and let Claude treat its instructions as a system prompt?
Start at the official repo
The canonical repo, github.com/anthropics/skills, is the gold standard. As of May 2026 it has roughly 134k stars and 15.8k forks, enough activity to be the de facto reference implementation. If you want to learn what a high-quality SKILL.md looks like, read this one first. Then read it again after you've written your own and want to know why yours is twice as long.
The repo contains around seventeen first-party skills bucketed into creative, technical, document, and enterprise. A few worth pulling open even if you never install them.
skills/pdf is a masterclass in what a "wraps a Python library" skill looks like when written by people who actually use it. Its description begins "Use this skill whenever the user wants to do anything with PDF files," and then enumerates the verbs (extract, merge, split, rotate, watermark, fill, encrypt, OCR) so the model has the trigger phrases it needs to auto-invoke. The skill leans on pypdf, pdfplumber, reportlab, and the pdftotext and qpdf command-line tools. The directory itself is what most third-party skills lack: a working SKILL.md, supporting reference docs, example scripts.
skills/skill-creator is the meta-skill. It walks you through capturing intent, drafting a SKILL.md, writing test cases, running evals with and without the skill loaded, and tuning the description until the model triggers it when you want it to. Read this one if you read nothing else. It's also the file most people skip and then regret skipping.
skills/webapp-testing and skills/mcp-builder round out the "actually does work" examples. The first drives headless browser automation; the second scaffolds a Model Context Protocol server, now the most common bridge between Claude and external APIs.
What's missing from the official repo is breadth. Seventeen skills cover the patterns. They do not cover your job. No skill for reviewing a Stripe webhook handler, no skill for drafting an OKR doc in your company's format, no skill for migrating a Rails 6 app to 7. The official repo is a textbook, not a library. The companion piece 15 real Claude skills you can install today walks through the working library that ships alongside it.
The wild long tail
Step outside the official repo and the picture changes fast. GitHub's code search for SKILL.md (authenticated, since unauthenticated searches now bounce you to a sign-in page) returns tens of thousands of files. SkillsMP, which aggregates skills from public GitHub repos, claims around 1.3 million indexed skills. That number is inflated by forks and by repos where every file is generated, but the order of magnitude is real. People publish skills faster than anyone can read them.
Run a SKILL.md search and you get a handful of well-maintained repos with genuine taste behind them, forks of those repos, forks of forks, repos where someone pasted a CLAUDE.md and renamed it, skills generated by an LLM in one shot with no test or eval and a description so vague the model never invokes it, skills that worked in October 2025 against a now-deprecated model, and one person's "deploy to AWS" skill that hardcodes their account ID. None of this is malicious. Most of it is sincere. The format makes it trivial to ship something, so discovery cost is high.
The aggregator-style repos worth a skim, all real:
github.com/ComposioHQ/awesome-claude-skills bills itself as a curated list of 1,000+ production-ready skills and plugins. Curated here means "a maintainer made a choice," not "every entry was tested last week." Still a usable starting index.
github.com/travisvn/awesome-claude-skills is a smaller awesome-list with a tighter selection, last updated April 2026. The kind of list you read top to bottom.
github.com/obra/superpowers is Jesse Vincent's fourteen-skill framework (Vincent and the rest of Prime Radiant) focused on agent development methodology: test-driven-development, systematic-debugging, dispatching-parallel-agents, using-git-worktrees. The repo is at v5.1.0 (May 2026), rare in this space because most skill repos never get versioned at all.
github.com/alirezarezvani/claude-skills collects 268 skills aimed at coding agents broadly (Claude Code, Codex, Gemini CLI, Cursor, and eight more), interesting more as a cross-tool index than a quality bar.
These help. They're still GitHub. The maintainer is one person, the freshness signal is whatever commit is on top, and the quality bar is "the list owner thought this was good enough to include."
How to evaluate a third-party skill before installing it
Installing a skill is closer to installing a shell script than a package. The allowed-tools frontmatter can grant Claude permission to run things like Bash(git push *) or Bash(rm *) without prompting, per the docs. If you trust the repo, fine. If you don't, read the skill before it reads you.
A short checklist that catches most of the bad ones.
Open the SKILL.md and read the frontmatter. A real skill has a description that names trigger phrases ("Use when the user asks for X" or "Use whenever a .pdf file is mentioned"). A bad one says "Helps with code" and will either never trigger or trigger on everything.
Check allowed-tools. If it lists Bash(*) or hands out Edit plus Bash(git push *) plus Bash(rm *) with no narrowing, you're giving up the keys. Fine for a skill you wrote, a stop-and-think for a stranger's repo.
Look at the supporting files. A serious skill ships scripts, reference docs, examples. A lone SKILL.md with five vague bullet points is usually a first attempt the author pushed and never iterated on.
Check the last commit date. The skill format has shifted at least twice: custom commands were merged into skills, dynamic shell injection landed, disable-model-invocation was added. A skill untouched since 2024 probably uses syntax that no longer means what it used to.
Look at the README. If there's no README explaining what installing this skill will let Claude do, the author hasn't thought about distribution. The skill might still be fine, but you're the QA.
Check who wrote it. A skill from someone you can find on the public internet (a known engineer's GitHub, a company org account, a maintainer of something you already use) is a different proposition from an account with zero followers and one repo. Neither auto-disqualifies; both are signal.
Run grep -r on the skill folder for absolute paths, API keys, account numbers, and hardcoded URLs. People paste their own environments into skills more often than they should.
This is roughly fifteen minutes per skill done right. Multiply by however many you're evaluating and the time-to-find-a-good-skill becomes the most expensive part of the workflow.
The curated alternatives
Several projects have noticed this and built a Claude skills directory on top of GitHub instead of asking you to crawl it.
Claude Marketplaces ranks skills, MCP servers, and plugin marketplaces by install count, GitHub stars, and community votes. It lists around 4,200 skills, 770 MCP servers, and 2,500 marketplaces across thirty-plus categories. Filtering by category and sorting by installs gets you to the top 1% of a category in about three clicks. GitHub won't do that for you. The Claude Skills Marketplace piece walks through the fragmentation in more detail.
SkillsMP goes the other direction. It aggregates rather than curates: about 1.3 million skills pulled from public repos, a minimum-two-stars filter, semantic search across descriptions. Useful if you want breadth and don't mind doing your own quality pass. Its FAQ tells you to treat community skills like any open-source code and inspect before use, which is the right framing.
Knack takes a different bet. Instead of cataloging every skill on GitHub, we keep a smaller hand-reviewed library and let you author your own through an interview instead of by editing YAML. Each library skill is tested against real prompts before shipping, the frontmatter is normalized, and the install path is knack pull <slug> instead of git clone followed by an audit. Best for people who want a working skill in five minutes more than they want to read someone else's SKILL.md. Worth checking against SkillsMP and Claude Marketplaces; the right tool depends on whether you'd rather browse a million entries or a thousand, and how much you want to write skills yourself.
None of these solve the underlying issue: the format is too easy to publish to. The same property that makes skills delightful to write makes them painful to find. A directory is a partial fix. Reputation, install counts, and update recency help. Eventually one of the directories becomes the npm of skills, and people write skills primarily for whichever registry runs the leaderboard. Until then, expect fragmentation.
A take
The skill format is going to win. The format is right, the timing is right, and the open spec means tools beyond Claude Code (Codex, the open Agent SDK, plugins) speak the same dialect. None of that changes the fact that finding a good skill on GitHub today is harder than writing one yourself, which is a symptom of how new the format is.
If you like archaeology, dig through anthropics/skills and obra/superpowers and read them like source code. You'll learn more about how to write a skill than any tutorial will teach you.
If you want a skill that works tomorrow morning, pick a curated source and stop browsing.
If the skill you found doesn't ship with an eval and a test prompt that demonstrates it triggering correctly, the author hasn't finished it. That's the bar.
The skill registries that survive will be the ones that show you, in one screen, whether a skill works against the current model, what tools it grants, who wrote it, and when. Everything else is a wiki.