1,000 skills is a landfill. This is a payroll of 44.
Every other list races to the biggest number. This one is a hiring decision: 44 agent skills, each with a job description, each link verified by hand. Clone one with ./hire.sh and it lands in .claude/skills/.
./hire.sh obra/superpowers DietrichGebert/ponytailCompatible with Claude Code · Cursor · Codex · Gemini CLI · OpenCode — anything that reads SKILL.md.
Built on the format Anthropic publishes for Agent Skills.
Every link verified 2026-07-23. Star counts are real, pulled from the GitHub API on that date — not copied from a screenshot. Run ./scripts/check-links.sh to re-verify the whole roster yourself.
The big lists have stopped being lists. The largest one has 1,077 open issues and hasn't been touched since May 22, 2026 — two months of dead links nobody swept. It also opens with a vendor banner and a UTM-tagged section for the vendor's own product. You came for an index and walked into a funnel.
A thousand entries is not curation. It is a landfill with a table of contents.
So this repo inverts it. Hundreds looked at, forty-four kept. Each one gets the thing a category label never gives you — when to hire it. No vendor sections. No affiliate links. No UTM parameters. No sponsor tier. If an entry is here, it is because it earns a seat.
⏸ = solid, but no commits in the last 30 days. Still worth hiring, just not actively moving.
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| superpowers | 259,621 | MIT | Agentic skills framework and development methodology | You want one opinionated loadout instead of assembling forty |
| ECC | 232,324 | MIT | Harness optimization: skills, instincts, memory, security | Your agent works but you can't tell why it sometimes doesn't |
| skills | 183,124 | MIT | Matt Pocock's .agents directory, TypeScript-first |
You write TS and want taste baked in, not explained |
| agent-skills | 79,937 | MIT | Production-grade engineering skills | You ship to real users and "it compiles" isn't the bar |
| ponytail | 88,037 | MIT | Makes the agent think like the laziest senior in the room | Your agent pulls a library in to format a date |
| get-shit-done | 64,783 | MIT | Meta-prompting and context engineering system | You keep re-explaining the same context every session |
| planning-with-files | 25,644 | MIT | File-based planning with crash-proof recovery | Long runs die halfway and take the plan with them |
| skills | 163,527 | — | Anthropic's own Agent Skills repo | You want the reference format straight from the source |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| ui-ux-pro-max-skill | 109,113 | MIT | Design intelligence for UI/UX across platforms | Everything your agent builds looks like a Bootstrap demo |
| taste-skill | 66,553 | MIT | Stops generic, boring output; aesthetic judgment | The output is correct and completely forgettable |
| huashu-design | 21,838 | MIT | HTML-native prototypes, slides, animation, MP4 export | You need a clickable prototype, not a description of one |
| hallmark ⏸ | 15,821 | MIT | Anti-AI-slop design skill for Claude Code, Cursor, Codex | You can smell the AI in your own landing page |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| marketingskills | 41,284 | MIT | CRO, copywriting, SEO, analytics, growth engineering | You built the thing and nobody arrived |
| claude-seo | 12,102 | MIT | 25 sub-skills, 18 sub-agents: technical SEO, schema, local | You need SEO done as a pipeline, not as advice |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| caveman | 92,176 | MIT | Cuts ~65% of tokens by talking like a caveman | Your bill grew faster than your codebase |
| claude-mem | 88,294 | Apache-2.0 | Persistent context across sessions, compressed and injected | Every new session starts with you re-introducing yourself |
| rtk | 72,692 | Apache-2.0 | CLI proxy cutting 60–90% of tokens on common dev commands | git diff alone is eating your context window |
| mem0 | 61,509 | Apache-2.0 | Universal memory layer; ships as an official Claude connector | You want a real memory store, not a markdown file |
| context7 | 59,617 | MIT | Up-to-date library docs piped to the model | Your agent confidently uses an API that was removed in v3 |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| graphify | 94,102 | Apache-2.0 | Turns a codebase, SQL schemas, configs, PDFs into a queryable graph | Grep stopped scaling three refactors ago |
| Understand-Anything | 75,718 | MIT | Code into interactive, explorable knowledge graphs | You inherited a repo and nobody left a map |
| qmd ⏸ | 28,172 | MIT | Local BM25 + vector + LLM re-rank search over your docs | Your knowledge base is local and must stay local |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| scientific-agent-skills | 31,518 | MIT | 148 skills across biology, chemistry, drug discovery | The work is science, not CRUD |
| ARIS | 13,737 | MIT | Markdown-only skills for autonomous ML research loops | You want experiments running while you sleep |
| AI-Research-SKILLs | 11,006 | MIT | Research and engineering skills library, model-agnostic | You need the agent to read papers, not just code |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| Anthropic-Cybersecurity-Skills | 26,385 | Apache-2.0 | 817 skills mapped to MITRE ATT&CK, NIST CSF 2.0 | Your review needs to name a framework, not a vibe |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| agency-agents | 135,984 | MIT | A full agency: eng, design, growth, PM, QA, legal | One agent is a person; you need a company |
| gstack | 123,784 | MIT | Garry Tan's exact setup: CEO, Designer, Eng Manager, QA | You want a proven loadout, not a blank config |
| ruflo | 65,604 | MIT | Meta-harness for multi-agent swarms and autonomous flows | Coordination is now the hard part, not the work |
| agents | 38,162 | MIT | Multi-harness plugin marketplace across six agent CLIs | Your team is not all on the same tool |
| oh-my-claudecode | 37,990 | MIT | Teams-first multi-agent orchestration, zero learning curve | You want orchestration without reading a manual |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| learn-claude-code | 71,996 | MIT | A nano agent harness built from fundamentals | You want to understand the loop, not just run it |
| claude-code-best-practice | 63,331 | MIT | Vibe coding → agentic engineering, staged | You're past prompting and don't know what's next |
| BMAD-METHOD | 50,994 | see repo | 12+ agents: brainstorm → PRD → architecture → shipped | You need a process, not a smarter model |
| harness-engineering | 2,212 | CC-BY-4.0 | Anthology, field guide, agent context bundle | You reference docs from CLAUDE.md and want good ones |
| sample-specship | 208 | MIT | recon → plan → build → validate → ship, adversarial gates | Small repo, real idea: the builder can't grade itself |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| cc-switch | 120,307 | MIT | Desktop switcher across Claude Code, Codex, OpenCode, Grok | You run more than one agent CLI and hate the juggling |
| awesome-mcp-servers | 91,212 | MIT | The de-facto MCP server directory | Skills cover thinking; MCP covers touching real tools |
| obsidian-skills ⏸ | 43,037 | MIT | Obsidian CLI, Markdown, Bases, JSON Canvas | Your second brain should be writable by the agent |
| knowledge-work-plugins | 22,967 | Apache-2.0 | Anthropic's plugins for knowledge work in Claude Cowork | The job function isn't engineering |
| skills | 10,855 | Apache-2.0 | Drive the Hugging Face ecosystem from the terminal | Models, datasets and Spaces are part of your loop |
| skills ⏸ | 5,639 | MIT | Anthony Fu's curated personal collection | You trust taste over volume |
| hire | ⭐ | license | what it does | hire when |
|---|---|---|---|---|
| Skill_Seekers | 14,531 | MIT | Turns docs sites, repos and PDFs into skills, with conflict detection | The skill you need doesn't exist yet |
| baoyu-skills | 24,033 | MIT | Skills to write, illustrate and publish articles | The output is content, and it has to read like a human |
hire.sh clones a skill into .claude/skills/ and strips its .git so your repo stays clean.
# one hire
./hire.sh obra/superpowers
# several at once
./hire.sh DietrichGebert/ponytail Leonxlnx/taste-skill JuliusBrussee/caveman
# somewhere else
SKILLS_DIR=~/.claude/skills ./hire.sh addyosmani/agent-skillsDon't take the numbers on faith — that is the whole complaint about the other lists.
./scripts/check-links.shWalks every GitHub link in this README and reports anything that no longer resolves. data/skills.json holds the same roster machine-readable, with the star counts and the date they were pulled.
One rule: a pull request must argue for a seat, not for an entry. Say what the skill does and when you'd hire it. See CONTRIBUTING.md.
MIT. Built to be forked.