multi-agent-skills-catalog

# Multi-Agent Skills Catalog

[English](README.md) | [简体中文](README.zh-CN.md)

![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)
![Installable Skills](https://img.shields.io/badge/installable_skills-120-blue)
![Public Default](https://img.shields.io/badge/public_default-72-6f42c1)
![Curation](https://img.shields.io/badge/curation-Source%2BOverlap%2BContent-orange)
![Rankings](https://img.shields.io/badge/rankings-skills.sh%20primary%20%7C%20Context7%20secondary-black)

A curated distribution layer for multi-agent skills: use `find-skills` for open-ended exploration, and use this repo for profiles, deterministic reinstalls, and cross-agent sync.

Live site (GitHub Pages): https://louislau-art.github.io/multi-agent-skills-catalog/

Current snapshot: **120 installable public skills** in the catalog. The default public install profile (`public-default`) currently resolves to **69 skills**. A full `all-public` install yields the complete 120-skill public catalog (plus internal `.system`).

## Positioning

This project is intentionally **not** trying to replace raw skill discovery.

- `find-skills` is the right tool when you want to explore the long tail, search broadly, or pull in something ad hoc.
- this repo is the right tool when you want a curated manifest, scenario-based profiles, reproducible installs, post-install validation, and consistent cross-agent sync.

In short:

- `find-skills` = exploration
- this repo = curation, profiles, and reproducible installs

This repository intentionally contains:
- `skills_manifest.csv` (selected skills with source/score)
- `skills_selected.txt` (plain list)
- `profiles/` (public install bundles such as `core-meta`, `development-core`, `writing-blog`)
- `scripts/install_curated.py` (cross-platform one-click installer)
- `scripts/install_curated.sh` (thin Unix wrapper around the Python installer)
- `scripts/install_curated.ps1` (thin PowerShell wrapper around the Python installer)
- `scripts/validate_skills_frontmatter.py` (post-install validator/sanitizer for `SKILL.md` frontmatter)
- `global-context/` (tracked runtime source of truth shared by Codex / Claude / Gemini)
- docs for de-dup policy and stack classification

It intentionally does **not** contain third-party `SKILL.md` contents.

## Why this approach

- keeps discovery and curation as separate layers instead of mixing them
- lighter and easier to maintain
- deterministic reinstall from source
- avoids re-hosting third-party skill files
- easier to sync the same curated pack across multiple agent directories

## Quick Start

```bash
# install the default public profile to the Claude-compatible base target
python scripts/install_curated.py claude --profiles public-default

# install the full public catalog
python scripts/install_curated.py claude --profiles all-public

# install once, then sync to Codex + Gemini + CodeBuddy
python scripts/install_curated.py all --profiles public-default

# add writing/blog support on top of the default profile
python scripts/install_curated.py claude --profiles public-default+writing-blog

# job-search/resume focused install
python scripts/install_curated.py codex --profiles resume-job-search

# Unix convenience wrapper
bash scripts/install_curated.sh all --profiles public-default

# Windows PowerShell wrapper
powershell -ExecutionPolicy Bypass -File .\scripts\install_curated.ps1 all --profiles public-default

# inspect available public profiles
python scripts/install_curated.py --list-profiles

# dry-run first
DRY_RUN=1 python scripts/install_curated.py claude --profiles public-default+cloud-platform
```

PowerShell dry-run example:

```powershell
$env:DRY_RUN = "1"
.\scripts\install_curated.ps1 qwen --profiles public-default
```

## Public Profiles

This repo now distinguishes:

- **public catalog**: all rows in `skills_manifest.csv`
- **public profiles**: install bundles under `profiles/`
- **private local overlay**: local-only additions kept outside the public product surface

Think of the rankings and `find-skills` ecosystem as the discovery surface, and this repo's manifest + profiles as the install surface.

Current public profiles:

- `core-meta` — discovery, verification, review, planning, session continuity
- `development-core` — software development starter pack
- `context7-integration` — Context7 MCP plus reusable docs lookup workflow
- `writing-blog` — public writing/blogging starter
- `resume-job-search` — resume and job-search starter
- `docs-office` — PDF/DOCX/PPTX/office workflows
- `cloud-platform` — Vercel/Supabase/Hugging Face/platform workflows
- `design-ui` — design and frontend-heavy work
- `database-data` — database, RAG, and data workflows

Installer aliases:

- `public-default = core-meta + development-core`
- `all-public = union of all public profiles`

Context7 setup guide:

- `docs/context7-agent-setup.md`
- `docs/context7-agent-setup.zh-CN.md`

Supported targets:
- `claude` (default)
- `codex` (install via Claude-compatible target, then sync to `~/.codex/skills`)
- `gemini` (sync to `~/.gemini/skills`)
- `qwen` (alias of `gemini`; uses the same skills directory)
- `codebuddy` (sync to `~/.codebuddy/skills`)
- `all` / `claude+codex+gemini+codebuddy`
- custom combos such as `claude+codex`, `claude+gemini+codebuddy`, `claude+qwen`
- `universal`, `global`, `cursor`, `auto` (install-only targets; no post-install sync)

The installer reads `skills_manifest.csv` as the public catalog, resolves one or more profile files from `profiles/`, installs the matching upstream skills, validates/sanitizes known `SKILL.md` frontmatter issues in the Claude-compatible base install, then copies the resulting local skill tree into compatible agent directories. It does **not** vendor third-party `SKILL.md` files into this repo.

### Directory Overrides

If your local agent uses a non-default path, set an override before running the installer:

```bash
export CLAUDE_SKILLS_DIR=/custom/claude/skills
export CODEX_SKILLS_DIR=/custom/codex/skills
export GEMINI_SKILLS_DIR=/custom/gemini/skills
export CODEBUDDY_SKILLS_DIR=/custom/codebuddy/skills
```

`qwen` reuses `GEMINI_SKILLS_DIR`.

### Codex-First Sync

If you use `~/.codex/skills` as the primary user skills directory, sync other agents from Codex instead of treating `.claude/skills` as the source of truth:

```bash
# Preview what would change
python scripts/sync_from_codex.py --dry-run --prune

# Copy Codex user skills into Claude/Gemini/CodeBuddy
python scripts/sync_from_codex.py --prune

# Or mirror each skill directory as symlinks
python scripts/sync_from_codex.py --mode symlink --prune
```

This script syncs user skill directories only and leaves target-side `.system` folders untouched.

## Three-Agent Runtime Sync

Use the tracked repo source of truth when you want skills, global context, and MCP to stay aligned across Codex / Claude / Gemini:

```bash
python scripts/sync_agent_context.py --mode symlink
python scripts/sync_mcp.py
```

What this does:

- mirrors `global-context/AGENTS.md` into each runtime-specific global instruction file
- ensures OpenCode loads the shared file through its `instructions` config
- syncs managed MCP servers from `global-context/mcp-servers.json`

## Files

- `skills_manifest.csv`: `slug, skill_name, source, installs, trust, score`
- `skills_selected.txt`: compatibility export of the current public catalog
- `profiles/*.txt`: public install profiles
- `manifest_summary.json`: generation metadata
- `scripts/install_curated.py`: cross-platform installer and sync entry point
- `scripts/install_curated.sh`: Unix wrapper for the Python installer
- `scripts/install_curated.ps1`: PowerShell wrapper for the Python installer
- `scripts/validate_skills_frontmatter.py`: validate and sanitize known frontmatter issues in installed skills
- `scripts/fetch_skills_sh_rankings.py`: pull the prerendered `skills.sh` leaderboard for the static site
- `scripts/fetch_context7_skill_rankings.py`: pull live ranked skills from Context7 A

[truncated…]