[!TIP]
Autonomous by design. Point any agent at this repo and it selects the
right skill or prompt for each task on its own: it reads AGENTS.md
(and CLAUDE.md in Claude Code), then picks from
index.json by each entry's use_when trigger. You install
once; you never have to name a skill.

What this is

One place that finds every skill. Two halves make that true:

• The library: original skills and prompts, written for this repository,
  every one complete and tested. Below.
• The ecosystem index: every major skill collection on
  GitHub, indexed with per-skill links for the official ones. 8,000+ skills
  reachable through ecosystem.json, one fetch away.

The idea behind the library: a tested method beats improvisation. Hand an
agent the way a strong engineer reviews code, and its review improves in one
step. Hand a person a prompt built on what actually works, and the answer
improves on the first try.

• skills/ are working methods an agent loads and follows:
  reviewing code, debugging, writing a postmortem, designing an API. One
  folder per skill, one SKILL.md inside, in the format Claude agents load
  natively and every other tool can read.
• prompts/ are complete prompts with named {variables},
  the model settings to run them, and one honest line on what each does well.
  Build your own in the browser with the
  Prompt Studio.
  Where settings mention temperature: that is the model's freedom to
  improvise, 0 exact and repeatable, 1 creative. No temperature control in
  your tool? Skip it; the prompt works at the default.
  In Claude Code every prompt installs as a slash command: /summarize,
  /tldr, /explain, /plan, /brainstorm, /critique, /improve,
  /outline, /steps, /pros-cons, plus /goal, /autoresearch, and
  /reflect (the generated commands/ folder). Cursor and
  Antigravity read the same prompts as slash commands too;
  INSTALL.md has the one-line copy step per tool.

Each entry is plain markdown with a short header. That is the design, not a
limitation: a method an agent can read is one you can read, edit, version, and
carry to your next tool. No runtime, no framework, no format that expires when
a product does.

Quick start

Claude Code adds the whole library from one command, organized as one
installable plugin per category:

/plugin marketplace add Amey-Thakur/AI-SKILLS

Any other tool reads plain markdown. The one-liners:

<raw> above is
https://raw.githubusercontent.com/Amey-Thakur/AI-SKILLS/main. Full per-tool
instructions, including scoped Cursor .mdc rules and Copilot instruction
files, are in INSTALL.md.

For agents: the whole catalog is machine-readable at
index.json (each entry with a description and a raw URL),
mirrored in llms.txt, with usage rules in AGENTS.md.
Fetch the index, pick by description, pull only what the task needs.

What's inside

One row per category; every entry, with its one-line description, is in
CATALOG.md.

Principles

1. Portable. Plain markdown + minimal YAML. If a tool dies, the content
   survives.
2. Honest. Each entry says what it is for and where it does not apply.
   Nothing here claims to replace judgment.
3. Complete. An entry ships when it is usable end to end, not before.
4. Small. One method per skill, one job per prompt. Composition beats
   bloat.

FAQ

Why not just prompt harder? "Review this well" leaves the agent to guess
what good means. A skill hands it the priorities, the verification steps, and
the reporting format a strong engineer would use. The output changes on the
next run.

Does this work with my tool? If the tool reads a markdown file, yes. The
per-tool steps in INSTALL.md are conveniences, not requirements.

How is this different from a curated list? It is both halves. The library
is the things themselves: every entry here, complete, in one format. The
ecosystem index is the list: every major collection elsewhere,
linked and machine-readable. One fetch covers both.

Can I use these commercially? Yes, under MIT. Attribution is appreciated
and never required.

Contributing

New skills and prompts are welcome when they clear the bar in
CONTRIBUTING.md: complete, self-contained, honest,
distinct, portable. By taking part you agree to the
Code of Conduct. Found something that could cause harm?
Follow the security policy, not a public issue.

License and author

Released under the MIT License. Use these skills and prompts
commercially or privately; attribution is appreciated and never required.

Built by Amey Thakur, who also builds
NotebookLab, the offline-first AI
knowledge workspace these methods grew out of.