One command installs curated, battle-tested Agent Skills that slash token usage of AI coding agents — without lowering output quality.
Quick start · Why it saves tokens · Supported agents · Catalog · FAQ · فارسی
Every AI coding agent — Claude Code, OpenAI Codex, Google Antigravity, Cursor, GitHub Copilot, Gemini CLI — burns tokens on the same three things:
- Re-discovery. Each session, the agent re-explores your repo to learn conventions it already learned yesterday.
- Bulky always-on context. Giant rule files (
AGENTS.md,CLAUDE.md,.cursorrules) get resent on every single turn. - Trial and error. Vague tasks → exploration → wrong turns → fix-up rounds. Each round resends the whole conversation.
You feel it as: hitting context limits, slow turns, and a token bill that keeps growing.
Agent Skills are the open standard (agentskills.io) adopted by 40+ clients — including Claude Code, Codex, Copilot, Cursor, Antigravity and Gemini CLI. A skill is a folder with a SKILL.md; agents load them progressively:
| Mechanism | What happens | Token effect |
|---|---|---|
| Progressive disclosure | Only each skill's name + description (~50–100 tokens) sits in context; full instructions load only when a task matches | Replaces bulky always-on rule files |
| Scripts instead of context | Deterministic work (documents, analysis, validation) runs as code, not LLM reasoning | That work costs zero LLM tokens |
| Battle-tested procedures | Curated workflows remove exploration, wrong turns and fix-up rounds | Fewer turns, better results |
| Cache-friendly design | Stable prefixes + lean sessions → more prompt-cache hits | Cached input tokens are discounted up to ~90% by major providers |
Quality goes up, not down: skills encode verified procedures instead of the agent improvising a new approach every session.
git clone https://github.com/tmolavi/lean-agent-skills.git
cd your-project
../lean-agent-skills/install.sh # project-level (recommended)
# or globally, for every project:
../lean-agent-skills/install.sh --global
# preview first:
../lean-agent-skills/install.sh --dry-runThat's it. Start a new turn in your agent — skills are discovered automatically, or invoke them explicitly (e.g. $skill-name in Codex, /skills listing).
Options
--dest DIR base directory for the project install (default: current dir)
--global install into your home directories instead
--agents LIST codex,claude,antigravity,cursor,copilot,gemini
(default: codex,claude,antigravity,copilot)
--dry-run show what would happen
--force overwrite an existing AGENTS.md
The installer also writes a tiny token-frugal AGENTS.md (~20 lines) with working rules for context hygiene, scope discipline and targeted execution — and never overwrites yours without --force.
Remove everything with uninstall.sh (it tracks exactly what it installed).
Full breakdown with skill paths and official docs: docs/AI-CODING-AGENTS.md
| Agent | Vendor | Skill path (project) | Status |
|---|---|---|---|
| Claude Code | Anthropic | .claude/skills/ |
✅ native |
| Codex | OpenAI | .agents/skills/ |
✅ native |
| Antigravity | .agents/skills/ |
✅ native | |
| Cursor | Anysphere | .cursor/skills/ |
✅ native |
| GitHub Copilot | GitHub/Microsoft | .github/skills/ |
✅ native |
| Gemini CLI | .gemini/skills/ |
✅ native | |
| Windsurf | Codeium | .windsurf/skills/ |
✅ native |
| OpenCode | community | .opencode/skills/ |
✅ native |
| Zed, Kiro, Junie, Amp, Goose, Roo Code, Cline, Aider… | various | standard-compliant | ✅ via agentskills.io |
This repo selects the best existing skills from official sources — no AI-slop, no bulk generation. 45+ skills, all lazy-loaded:
| Source | Skills | Why it's here |
|---|---|---|
| anthropics/skills (official) | skill-creator, docx, pdf, pptx, xlsx |
Document work delegated to scripts (zero LLM tokens); skill-creator helps you build your own |
| openai/skills (official Codex catalog) | 30+ curated: gh-fix-ci, gh-address-comments, define-goal, security-best-practices, playwright, deploys (Vercel/Netlify/Cloudflare/Render), Notion, Linear, Figma… |
OpenAI's vetted workflows replace trial-and-error |
| multica-ai/andrej-karpathy-skills | karpathy-guidelines |
"Surgical changes, simplicity, verifiable success criteria" — the quality guardrail |
| hqhq1025/skill-optimizer | skill-miner, skill-generalizer, skill-personalizer |
Mines your session history into new skills, then keeps every skill lean |
Browse even more: VoltAgent/awesome-agent-skills · gmh5225/awesome-skills · agent-skill.co
The machine-readable list lives in manifest.yaml.
## Context hygiene
- New task = new session; compact/summarize when context grows heavy.
- Read only the files you need; prefer search/grep over reading whole files.
- Never re-read a file that has not changed since you last read it.
## Scope discipline
- Make minimal, surgical changes; do not touch files outside the task.
- For anything beyond a small edit: present a short plan first, then execute.
## Execution
- Run targeted tests, not the whole suite, unless asked.
- If an installed skill covers the task, use it and follow its scripts.
- Keep diffs, comments and answers concise; no boilerplate.Deep dive with sources: docs/WHY.md · FAQ: docs/FAQ.md
Does this actually reduce my token bill? The mechanisms are structural: leaner always-on context, lazy-loaded instructions, scripted deterministic work, fewer exploration/fix-up turns, and better prompt-cache hit rates. Your exact savings depend on workload — but every mechanism above removes tokens that were previously spent on every turn.
Will output quality drop?
The opposite is the goal: curated skills replace improvisation with verified procedures. The included karpathy-guidelines skill explicitly enforces simplicity and surgical changes.
Do I need all 45 skills?
No. Skills are lazy-loaded; each costs only its name+description in context until used. Prune anytime — delete folders from the skills directory, or run skill-personalizer to tailor the set.
Which agents work? Anything implementing the open Agent Skills standard — Claude Code, Codex, Antigravity, Cursor, Copilot, Gemini CLI, Windsurf, OpenCode and ~40 more clients listed on agentskills.io.
More questions? → docs/FAQ.md
- Per-agent token-spend audit skill (uses local session logs)
-
npx lean-skillsone-liner installer - Community skill nominations via PR (quality bar: real usage, lean, model-agnostic)
- Quarterly re-curation against upstream repos
Propose a skill, report a broken upstream link, or improve the frugal rules. Quality bar is in CONTRIBUTING.md.
Part of the Molavi AI Engineering Ecosystem:
- mcp-agent-skills-hub: Curated production agent skills catalog.
- n8n-agent-skills: Production n8n agent workflow architecture.
- agent-project-discovery-skill: Codebase startup & context discovery skill.
- geo-scope: Multi-model empirical AI visibility benchmark engine.
- Ecosystem Map: Complete architecture and evidence flow.
Developed by Taghi Molavi — molavi.pro
MIT — see LICENSE. All skills belong to their respective upstream authors and licenses; this repository selects, organizes and installs them.