A skill is a folder with a SKILL.md file that your AI agent loads when a task matches it. Development skills encode process rather than taste: write the failing test first, reproduce before patching, check this before merging.
These twelve cover the full loop, from plan to ship. Star counts, licences and update dates come from GitHub, not from estimates.
| Your problem | Install | Why |
|---|---|---|
| I want one rule that fixes everything | tdd | A failing test is the only definition of done an agent cannot argue with. |
| The agent guesses at bug fixes | diagnosing-bugs | Gated phases stop it patching before it has a reproduction. |
| I want a full process, not one rule | superpowers | An entire methodology from brainstorm through to review. |
| My code handles user input | security-and-hardening | Agents ship working code that is quietly unsafe. |
Ordered by what to install first, not by star count. Stars and forks are repository totals, so skills sharing a repo share a number. All figures checked on 6 September 2026.
obra obra/superpowers
A complete development methodology for agents, not a single rule.
You want process across the whole job
Start here if you want opinionated structure end to end.
Matt Pocock mattpocock/skills
Forces the agent to write a failing test before it writes any code.
You want one rule that fixes everything
A failing test is the one definition of done an agent cannot argue with.
Matt Pocock mattpocock/skills
A gated loop that will not let the agent patch anything before it can reproduce it.
The agent is guessing at fixes
Guessing before reproducing is exactly where AI debugging goes wrong.
Matt Pocock mattpocock/skills
Reviews a branch on two separate axes and reports them side by side.
Before you merge anything
Separates this is ugly from this is wrong, and only one should block a merge.
Addy Osmani addyosmani/agent-skills
Hunts for the vulnerabilities an agent ships without noticing.
Your code touches user input
The gap almost every other skills list leaves wide open.
Addy Osmani addyosmani/agent-skills
Finds the real bottleneck across the whole stack, not just the bundle size.
Something is measurably slow
From the engineer who leads Chrome, and it covers the backend too.
Matt Pocock mattpocock/skills
Gives your agent a shared vocabulary for your project so its naming stops drifting.
Agents touch this codebase repeatedly
Naming drift is what makes long agent sessions quietly degrade.
Matt Pocock mattpocock/skills
Interviews you about a plan until every open decision is actually resolved.
Before any code exists
The cheapest bug fix is the one caught before you build anything.
Vercel vercel-labs/agent-skills
Vercel Engineering's own performance rules for React and Next.js.
You are building on React or Next.js
First-party guidance from the people who build the framework.
Anthropic anthropics/skills
Lets the agent drive a real browser so it can see its own runtime errors.
It compiles but nobody has seen it work
An agent that reads its own errors stops guessing why the page is blank.
Anthropic anthropics/skills
Builds MCP servers so your agent can reach systems it currently cannot touch.
The blocker is access, not reasoning
Skills change how an agent thinks; MCP changes what it can reach.
Anthropic anthropics/skills
Writes, tests and tunes skills of your own.
You want to build your own
A skill that never fires is worth nothing, however good its contents.
Start with one process skill
Install a single skill that governs how work gets done, either a full methodology like superpowers or one hard constraint like tdd. Not both on day one.
Copy the folder in
Clone the repo and copy the folder containing SKILL.md into .claude/skills in your project, or the same path in your home directory to use it across every project.
Add specialists on demand
Add diagnosing-bugs when the agent starts guessing at fixes, security-and-hardening before you handle user input, performance-optimization when something is measurably slow.
Read it before you trust it
A skill is instruction text your agent will follow. Read the SKILL.md, prefer named maintainers and recent commits, and be wary of anything fetching remote content.
Define done before it starts.
Left alone, an agent decides for itself when work is finished, and it is optimistic. A failing test, a reproduction step, a review axis: each one converts a vague request into a check that either passes or does not.
changes how it thinks
changes what it can touch
An open SKILL.md spec, portable across Claude Code, Cursor, Codex and more.
A test-first constraint. It is the one definition of done an agent cannot argue with.
A skill is instruction text your agent obeys. Read it before installing, like any dependency.
A skill is a folder containing a SKILL.md file, holding instructions an AI coding agent loads on demand when a task matches its description. Development skills encode process rather than taste: how to write a failing test first, how to reproduce a bug before patching it, what to check before merging. The agent reads them at the moment they become relevant instead of holding everything in context at once.
No, and installing all twelve at once is a mistake. Start with one process skill, either superpowers for a full methodology or tdd for a single constraint that improves everything. Add specialists when you hit the specific problem they solve: diagnosing-bugs when an agent starts guessing at fixes, security-and-hardening before you handle user input, performance-optimization when something is measurably slow.
Mostly yes. SKILL.md is an open format, so Cursor, Codex CLI, Gemini CLI and other harnesses read the same structure, and the written guidance is portable. What is not portable is tooling. Skills that spawn sub-agents, call harness-specific scripts, or rely on a particular plugin system will lose those parts elsewhere, though the instructions themselves still apply.
Copy the folder containing SKILL.md into .claude/skills in your project, or into the same path in your home directory to make it available across every project. Restart your agent so it registers. Several of these repos also ship as plugins with their own install command, so check the specific README, because the commands differ between repositories.
Because a test is the only unambiguous definition of done that an agent cannot talk itself out of. Without one, the agent decides for itself when the work is finished, and it is optimistic. A failing test written first converts a vague request into a binary check, which is why TDD skills consistently outperform more elaborate prompting strategies.
No, and they solve different problems. A skill changes how an agent thinks about a task, because it is instructions. An MCP server changes what an agent can touch, because it is tools and data access. If your agent produces poor quality work, you need a skill. If it cannot reach your database, your issue tracker, or your browser, you need an MCP server. Most serious setups run both.
A skill is executable instruction text that your agent will follow, so treat it like any dependency. Read the SKILL.md before installing it, prefer repositories with named maintainers and recent commits, and be cautious with any skill that instructs the agent to run scripts, fetch remote content, or modify files outside the project. Every repository on this page is public and named, so you can read the source before you trust it.
Every repository here is public and free to read. Licences differ, and each card shows the one its repository publishes, so check that before shipping a skill's contents inside a commercial product. A repository marked as having no licence means the author published no explicit terms, which means default copyright applies and you should ask before redistributing it.
Nothing on this page is ours. These are the repositories both lists draw from, so go star the ones you end up using.
Send the repo. If a link on this page has gone dead, tell us that too, because we would rather hear it from you than from Google.