Awesome Agent Skills

A curated, opinionated index of production-tested skills, plugins, and CLAUDE.md packs for AI coding agents — Claude Code, Codex, Cursor, OpenCode, Aider, Gemini CLI, and more.

This is not a directory dump. Every entry is hand-picked, briefly explained, and tagged with when to use it and when not to use it. Opinions are explicit, not hidden behind neutral descriptions.

If a skill below looks great but you've found it underwhelming in practice, open an issue — that's how this list stays useful.

Note on naming: The community uses "skills", "plugins", "personas", "rules", "instructions" interchangeably. They are all variations of: some markdown file the agent reads to change its behavior. This list collects the good ones, regardless of label.

---

Contents

• General Coding Behavior
• Language-Specific
• Code Review
• Testing
• Documentation
• Security
• Refactoring & Cleanup
• Workflow & Process
• Domain-Specific (Frontend, Backend, Data, etc.)
• Token Efficiency
• Frameworks & Build Systems
• Skill Authoring (meta)
• Contributing

---

General Coding Behavior

These shape how the agent thinks, communicates, and structures work — independent of language.

andrej-karpathy-skills 🌟

• Repo: multica-ai/andrej-karpathy-skills
• What it is: A single CLAUDE.md derived from Andrej Karpathy's observations on common LLM coding pitfalls (over-eager abstraction, premature generalization, etc.).
• When to use: As your default base CLAUDE.md for any project. Pairs well with more specific skills layered on top.
• When not: If you're already using a stricter, project-specific guide. Don't blend conflicting "voices."
• Why it works: Names anti-patterns the agent actually exhibits (rather than vague "write good code" instructions).

obra/superpowers 🌟

• Repo: obra/superpowers
• What it is: A full agentic-skills framework + opinionated software-development methodology.
• When to use: If you want a structured spec → plan → implement → verify workflow that the agent follows consistently.
• When not: Tiny scripts or one-off changes — the overhead isn't justified.
• Why it works: Treats the agent like a junior engineer who needs a process, not a vibe.

kiro_planner (built-in)

• Where: Bundled with Kiro CLI as the planner agent role.
• What it is: A planning-only persona that produces a step-by-step implementation plan before touching code.
• When to use: Multi-file changes where understanding scope first prevents redo loops.
• When not: Trivial fixes (typos, single-function edits) — straight-to-execution is faster.

---

Language-Specific

Python

• NVIDIA/skills — official agent skills published by NVIDIA, includes Python coding conventions and CUDA-aware practices.
• Imbad0202/academic-research-skills — research workflow (write → review → revise) tuned for Python data work.

TypeScript / JavaScript

• dotnet/skills — published by Microsoft .NET team but the patterns (strict typing, contract testing, docs-first) port well to TypeScript/Node.
• Looking for a great pure TS skill — PR welcome.

Rust

• Underserved category. If you maintain a tested Rust CLAUDE.md, please contribute.

Go

• Underserved category. If you maintain a tested Go CLAUDE.md, please contribute.

Multi-language

• Lum1104/Understand-Anything — turns a codebase into an interactive knowledge graph the agent can query. Works across most languages via tree-sitter.

---

Code Review

• colbymchenry/codegraph — pre-indexed code knowledge graph for Claude Code, Codex, Cursor, OpenCode. Cuts tool calls and tokens dramatically when the agent needs to navigate a codebase.

  Honest note: Adds a ~2k–5k-token graph header to every prompt. Worth it for repos > 10k LOC; overkill for tiny ones.

---

Testing

• Looking for a curated test-driven CLAUDE.md. The Anthropic Claude Code best practices doc covers TDD-with-agents but isn't packaged as a skill.

---

Documentation

• Imbad0202/academic-research-skills — academic writing pipeline (research → write → review → revise → finalize). Good template for any technical-writing skill.

---

Security

• Underserved category right after the May 2026 GitHub / Nx Console / Mini Shai-Hulud attacks. If you've packaged "do not exfiltrate", "verify dependencies", "treat .env files as untrusted" as a skill, please contribute.

• ATOM00blue/extensionguard — not a skill, but a CLI you can have your agent call before installing VS Code extensions.

• ATOM00blue/safeinstall — not a skill, but a CLI your agent can wrap around npm install.

---

Refactoring & Cleanup

• Looking for a "code archaeology" skill — one that has the agent investigate why something exists before refactoring it.

---

Workflow & Process

• obra/superpowers — see General Coding Behavior. The methodology layer of this is what makes it shine.
• anthropics/claude-plugins-official — Anthropic-managed directory of plugins for Claude Code. The official baseline.

---

Domain-Specific

Frontend / UI

• Underserved. If you have a "design-system aware" skill (knows your tokens, components, accessibility constraints), contribute.

Backend / API

• Underserved.

Data / ML

• Imbad0202/academic-research-skills — research → analysis → publication pipeline.

Game Development

• Empty. Open invitation.

---

Token Efficiency

The category that exploded in March-May 2026 as Anthropic and OpenAI tightened usage limits.

• drona23/claude-token-efficient — a single CLAUDE.md that keeps Claude responses terse on heavy workflows. Drop-in.
• colbymchenry/codegraph — pre-indexed knowledge graph (covered above). Reduces grep/read tool calls.
• ATOM00blue/codemap — (coming soon in this index series) universal single-file code map for any agent.

Anti-pattern alert: Don't combine multiple "be terse" skills — they conflict and confuse the agent. Pick one.

---

Frameworks & Build Systems

• Looking for: framework-specific skills (Next.js, Django, Rails, Bun, Vite). Contribute if you have one tested in production.

---

Skill Authoring (meta)

Resources for writing your own skills:

• Anthropic — Claude Code Best Practices — official guide. Read this first.
• obra/superpowers — methodology section — opinionated take on what a "skill" should contain.
• andrej-karpathy-skills — minimalist single-file approach worth studying.

Style guide

A skill is good if:

1. It names anti-patterns the agent actually exhibits. Vague "write clean code" advice is ignored.
2. It includes counter-examples. "Do this. Don't do that." beats abstract principles.
3. It is < 500 lines. Larger skills get the early sections forgotten.
4. It is project-agnostic. Project-specific stuff belongs in AGENTS.md / CLAUDE.md at the project root.
5. It declares scope. "This skill is for review, not for implementation" prevents misuse.

---

Anti-patterns (what NOT to do)

These keep showing up. Don't add them to this list.

• "You are an expert..." prompts that are 80% flattery, 20% content. Cut to the content.
• Massive numbered lists of 20+ rules. The agent silently drops the back half.
• Conflicting voices in the same skill. Pick one tone.
• Skills that try to do everything. Better: one skill per concern.
• Skills that hard-code a specific model. "Use Claude Opus 4.7" ages badly.
• Skills that disable safety behaviors. Hard pass. We don't list those here.

---

Contributing

Pull requests are very welcome. Quality bar:

• You've used the skill in a real project, not just read its README.
• You can describe when not to use it (every skill has a downside).
• It is open-source with a permissive license.
• No duplicates — search the existing list first.

To submit:

1. Open a PR adding the entry to the right section in this README.

2. Use the existing template:

   ### [skill-name](https://github.com/owner/repo)
   
   - **What it is:** _one sentence_
   - **When to use:** _real use case_
   - **When not:** _honest downside_
   - **Why it works:** _1-2 sentences_
   

3. If a category is empty (most are right now), highlight that you're filling it.

See CONTRIBUTING.md for the full details.

---

License

This list itself is CC0-1.0 (public domain — copy it, fork it, embed it). The linked skills retain their own licenses.

---

Why this list exists

The pattern: a viral agent skills repo lands on GitHub trending. Within a week, fifteen "awesome-X-skills" lists appear, each linking the same five popular skills, with no curation, no tagging, no opinion. They become useless after their first week.

This list aims to be the opposite:

• Hand-curated, not automated. Every entry was used, not just starred.
• Opinionated. "When not to use" is as important as "when to use."
• Honest about gaps. The empty sections (Rust, frontend, security) are flagged so contributors know where they're needed.
• Stays current. A weekly bot checks for dead links and stale repos.

---

Maintained by @ATOM00blue. Last meaningful update: 2026-05-22.