OpenSkill — Project-Aware Skill Discovery & Installer

A package-manager-style skill that reads your project, discovers relevant Claude skills from multiple sources, and installs them interactively.

Commands

Workflow

Step 1 — Read the Project

When /openskill is invoked:

1. Check for CLAUDE.md in the project root. Read it if present — it contains project context, stack, and existing skill references.
2. Scan the codebase if no CLAUDE.md exists or to supplement it:
   • Look at package.json, pyproject.toml, Cargo.toml, go.mod, composer.json, etc. to detect stack and dependencies
   • Check for framework indicators: next.config.*, vite.config.*, tailwind.config.*, Dockerfile, .github/workflows/, etc.
   • Note the primary languages, frameworks, and tooling in use
3. Detect the scope from the command argument (e.g., /openskill frontend → focus on frontend-relevant skills). If no scope given, infer from the full project read.

Summarize what you found in 2–3 sentences before proceeding.

Step 2 — Discover Skills

Search for relevant skills across three tiers in order:

Tier 1: Local Skills (check first)

First detect the environment: if /mnt/skills/ exists, you're in Claude.ai. Otherwise assume Claude Code.

Claude.ai — scan these paths:

• /mnt/skills/user/ — user-installed skills
• /mnt/skills/public/ — built-in public skills
• /mnt/skills/examples/ — example skills

Claude Code — scan these paths:

• ~/.claude/skills/ — user-installed skills
• .claude/skills/ — project-scoped skills (relative to project root)

For each skill found, read only its YAML frontmatter (name + description). Do not load full SKILL.md bodies — just names and descriptions are enough for matching.

Tier 2: Community GitHub (top starred repos)

Search GitHub for highly starred skill repositories. Good queries:

• claude skills topic:claude-skill
• SKILL.md claude anthropic
• claude-skills-registry

Reference references/github-search.md for how to evaluate and surface community skills.

Tier 3: Official Anthropic Skills

Check the official Anthropic skills repository if one exists. Search for anthropic/claude-skills or similar official registry. Reference references/official-registry.md for guidance.

Cross-check: After collecting candidates from all tiers, deduplicate by name. If a skill exists locally, prefer the local version — no need to suggest reinstalling it.

Step 3 — Match & Recommend

Match discovered skills to the project's stack and the requested scope. Present recommendations separated by source tier:

📦 Recommended Skills for [Project Name / Scope]

✅ Already installed:
  - [skill-name] — [one-line description]

📁 Local (built-in):
  1. [skill-name] — [one-line description]
  2. [skill-name] — [one-line description]

🌐 Community:
  1. [skill-name] (⭐ 342) — [one-line description]
  2. [skill-name] (⭐ 128) — [one-line description]

🏢 Official:
  1. [skill-name] — [one-line description]
  2. [skill-name] — [one-line description]

❓ Possibly relevant (lower confidence):
  - [skill-name] — [one-line description]

Max 5 per tier. Skip a tier entirely if nothing relevant is found — don't show empty sections.

Step 4 — Interactive Install

Ask the user which skills they want to install. Then ask:

"Install as user-scope (available in all your projects) or project-scope (this project only)?"

Detect the environment first: check if /mnt/skills/ exists — if yes, you're in Claude.ai. Otherwise assume Claude Code and use ~/.claude/skills/.

For each selected skill:

1. Copy or fetch the skill files to the chosen path
2. Confirm installation with: ✅ [skill-name] installed to [path]
3. If the skill has bundled resources (scripts/, references/, assets/), copy them too

After install, tell the user: "These skills will be available the next time Claude reads your project context."

Step 5 — Update CLAUDE.md (optional)

Offer to update or create the project's CLAUDE.md with a skills section:

## Installed Skills
- [skill-name]: [description]

Ask before writing. If the user says yes, append cleanly without overwriting existing content.

Scope Behavior

When a scope argument is passed (e.g., /openskill frontend), skip Step 1's full codebase scan. Instead:

• Use the scope as the primary filter for discovery
• Still check the project root for CLAUDE.md to get stack context
• Focus Tier 1/2/3 searches on that domain

Scope → keyword mapping:

• frontend → React, Vue, Svelte, CSS, Tailwind, UI, components, design systems
• backend → Node, Python, Go, Rust, Express, FastAPI, Django, databases, auth
• debug → logging, tracing, error handling, stack traces, breakpoints
• code-review → linting, formatting, best practices, refactoring, PR review
• api → REST, GraphQL, OpenAPI, Swagger, Postman, endpoints
• data → pandas, SQL, ETL, spreadsheets, analytics, charts
• infra → Docker, Kubernetes, Terraform, CI/CD, AWS, GCP, Azure
• testing → Jest, Pytest, Playwright, Cypress, unit/integration/e2e tests

Tone & UX

• Be concise. This is a CLI-style interaction — users expect fast, scannable output.
• Use emojis sparingly for status (✅ installed, 🆕 new, ❓ maybe, ⭐ community stars).
• Don't ask unnecessary clarifying questions. Read first, ask only if genuinely ambiguous.
• If nothing relevant is found, say so plainly and suggest the user try a different scope or check back later.

Reference Files

• references/github-search.md — How to find and evaluate community skills on GitHub
• references/official-registry.md — How to check the official Anthropic skills registry
• references/install-guide.md — Detailed install path logic and edge cases