The portable skill lifecycle manager for AI coding agents.
Search 300K+ skills across 6 registries. Import from GitHub in one command. Validate against the open spec. Remove cleanly with provenance tracking. Works with Claude Code, Codex, Cursor, Gemini CLI, and every platform that reads SKILL.md.
No runtime. No server. No lock-in. Your agent IS the engine.
The agent skills ecosystem is exploding — 300K+ skills across a dozen registries, marketplaces, and awesome-lists. But finding the right skill, installing it cleanly, and maintaining it across projects is still manual and fragmented.
skill-factory solves this with 9 slash commands, 1 agent, and 3 procedural skills that turn your agent into a self-service skill manager:
- Search across Skyll (89K+), LobeHub (100K+), skills.sh, openskills, and GitHub in a tiered cascade
- Import from any GitHub repo with
npxorgit clonefallback + provenance tracking - Validate against the agentskills.io open specification
- Audit quality — trigger descriptions, frontmatter, progressive disclosure
- Remove cleanly with manifest archival (never lose track of what was installed)
- Evaluate session-level skill effectiveness — what helped, what didn't, what was missing
/plugin marketplace add rooftop-Owl https://github.com/rooftop-Owl/rooftop-owl-marketplace.git
/plugin install skill-factoryInstalls all 9 commands + 1 agent + 3 skills with auto-discovery.
npx skills add rooftop-Owl/skill-factoryInstalls the 3 skills for any SKILL.md-compatible agent.
git clone https://github.com/rooftop-Owl/skill-factory.git
cp -r skill-factory/skills/* ~/.claude/skills/| Command | Description |
|---|---|
/skills-search <query> |
Tiered search across 6+ registries — Skyll API, LobeHub, skills.sh, openskills, GitHub topics, grep.app |
/skills-install <owner/repo> |
Import skills from GitHub with npx or git clone fallback. Tracks provenance in manifest |
/skills-remove <name> |
Remove an imported skill with manifest archival. Never lose history |
/skills-list |
Show all imported skills — source, status, install date |
/skills-create <name> |
Scaffold a new SKILL.md with valid frontmatter and directory structure |
/skills-validate <name> |
Validate against the agentskills.io spec — required fields, size limits, format |
/skills-audit <name> |
Quality audit — trigger description quality, progressive disclosure, body size |
/skills-health |
Health report across all installed skills — coverage, staleness, spec compliance |
/skills-eval |
Session-end skill performance evaluation — analyzes which skills helped, which didn't, and what was missed |
"import skill", "remove skill", "list imported", "validate skill"
Core lifecycle procedures for managing external skills. Handles the full import → use → maintain → remove cycle with manifest-based provenance tracking.
"create skill", "write SKILL.md", "skill frontmatter", "publish skill"
Authoring guide for writing production-quality skills. Covers the SKILL.md spec, description best practices (third-person, trigger phrases), progressive disclosure (body < 2K words, details in references/), and publishing to skills.sh.
"no skill match", "domain knowledge", "find skill", "acquire skill"
When you need a skill that doesn't exist yet. Three-tier acquisition: auto-install from trusted sources (MIT/Apache-2.0/MPL-2.0), manual review from spec sources, or synthesize from documentation.
At session end, run /skills-eval to find out whether the right skills were loaded and whether they actually helped.
The agent reads the current session transcript, identifies every skill that was loaded, and scores each as Effective, Neutral, or Ineffective. For underperformers, it diagnoses the root cause (wrong skill, outdated content, poor trigger, too broad, too narrow). For domains where no skill was loaded at all, it flags the missed opportunity.
Output is a structured report with per-skill findings and actionable next steps. Optionally triggers /skills-audit on flagged skills.
Requires Claude Code or OpenCode with session reading tools (
mcp_session_read).
skill-factory searches registries in a cascade — fast and high-signal sources first, broader and noisier sources as fallback.
| Tier | Source | Scale | Speed |
|---|---|---|---|
| T0 | Local inventory | Your installed skills | Instant |
| T1 | Skyll API + LobeHub | 190K+ | <2s |
| T2 | skills.sh, openskills, agent-skills-cli | 175K+ | 2-5s |
| T3 | GitHub topics + grep.app | All public repos | 3-8s |
| T4 | Curated awesome-lists | 2,000+ vetted | Reference |
| T5 | Anthropic official plugins | Curated | Separate path |
Stops at the first tier with 3+ relevant results. Full cascade documented in handbook/search-tiers.md.
skill-factory uses standard tools (npx, git, curl) and produces standard SKILL.md files. It works anywhere the open agent skills format is supported:
| Platform | Install Method |
|---|---|
| Claude Code | /plugin install skill-factory |
| Codex | npx skills add rooftop-Owl/skill-factory |
| Cursor | npx skills add rooftop-Owl/skill-factory |
| Gemini CLI | npx skills add rooftop-Owl/skill-factory |
| OpenCode | npx skills add rooftop-Owl/skill-factory |
| Windsurf | npx skills add rooftop-Owl/skill-factory |
| Roo Code | npx skills add rooftop-Owl/skill-factory |
| Any SKILL.md agent | git clone + copy to skills directory |
skill-factory/
├── .claude-plugin/
│ └── plugin.json # Claude Code plugin manifest
├── agents/
│ └── skill-evaluator.md # Session-end skill performance analysis
├── skills/
│ ├── skill-management/SKILL.md # Import, remove, list, validate
│ ├── skill-development/SKILL.md # Author and publish skills
│ └── external-skill-acquisition/
│ └── SKILL.md # 3-tier acquisition workflow
├── commands/
│ ├── skills-search.md # Tiered marketplace search
│ ├── skills-install.md # GitHub import with fallback
│ ├── skills-remove.md # Clean removal + archival
│ ├── skills-list.md # Inventory view
│ ├── skills-create.md # Scaffold new skill
│ ├── skills-validate.md # Spec validation
│ ├── skills-audit.md # Quality audit
│ ├── skills-health.md # Health report
│ └── skills-eval.md # Session-end evaluation
├── handbook/
│ ├── search-tiers.md # Full search cascade reference
│ ├── publishing-skills.md # How to publish to skills.sh
│ └── agentskills-io-spec.md # The open SKILL.md spec
├── CLAUDE.md # Plugin context for agents
├── LICENSE # MPL-2.0
└── README.md
Deep-dive references for skill authors and power users:
| Guide | Description |
|---|---|
| Search Tiers | Full probe cascade — 6 tiers, install commands, source maintenance notes |
| Publishing Skills | Step-by-step guide to publishing your skills on skills.sh |
| agentskills.io Spec | The open SKILL.md specification — frontmatter fields, format rules, validation criteria |
Agent IS the engine. skill-factory provides knowledge (skills), entry points (commands), and one analytical agent — your coding agent's native tool-calling handles all execution. No Python runtime, no server dependencies. Portable across platforms.
File-level copyleft (MPL-2.0). Modified files must stay open-source. Patent grant protects contributors. But you can freely combine skill-factory with proprietary projects in separate files. See LICENSE.
Tiered search over single-source. No single registry has everything. The cascade searches the fastest, highest-signal sources first and stops early when results are sufficient. This balances speed with coverage.
skill-factory is open to contributions under MPL-2.0:
- New registry backends — Add a search source to the tiered cascade
- Skill quality improvements — Better trigger descriptions, tighter procedures
- Platform adapters — Test and document installation on new agent platforms
- Handbook content — Guides for specific workflows or domains
Open an issue or PR at rooftop-Owl/skill-factory.
Mozilla Public License 2.0 — file-level copyleft with patent grant.
You can use, modify, and distribute freely. Modified files must remain open-source. New files you add can use any license. Patent claims on contributed code are automatically granted to all users.