Skill Finder

Discover and install agent skills from the skills.sh open ecosystem into Skill Compose.

When to Use This Skill

Activate when users:

Ask "how do I do X?" where an existing skill might help
Request "find a skill for X" or "is there a skill for X?"
Ask "can you do X?" for specialized tasks (poster design, data analysis, etc.)
Want to search for tools, templates, or workflows
Mention needing help with a specific domain that might have a community skill

Available Scripts

1. Search Skills — `find_skills.py`

Search the skills.sh ecosystem for skills matching a query.

python scripts/find_skills.py <query> [--limit N]

Example:

python scripts/find_skills.py "react performance"
python scripts/find_skills.py "docker" --limit 5

Output: JSON array of matching skills with name, source (owner/repo), installs count, and url (skills.sh link).

2. Install Skill — `add_skill.py`

Download a skill from GitHub and register it in Skill Compose.

python scripts/add_skill.py <owner/repo@skill-name>

Example:

python scripts/add_skill.py "vercel-labs/agent-skills@vercel-react-best-practices"

What it does:

Parses the owner/repo@skill-name identifier
Tries multiple GitHub paths to locate the skill (skills/<name>/, <name>/, root)
Downloads all skill files (SKILL.md, scripts/, references/, assets/)
Saves to the local skills/ directory
Registers the skill in Skill Compose via the import-local API
The skill is immediately available for use in Agent Presets

CRITICAL: Never Combine Questions with Tool Calls

When you ask the user a question or present results for them to review, your response MUST end with text only. Do NOT include any tool call (execute_code, bash, etc.) in the same response. The user needs a chance to read and reply. If you combine a question with a tool call, the tool executes immediately without waiting — this breaks the conversation flow.

WRONG (never do this):

"Should I install X?" + [execute_code: install X]

CORRECT (always do this):

Turn 1: "Should I install X?" (text only, no tool calls) Turn 2: User says "yes" Turn 3: [execute_code: install X]

How to Help Users Find and Install Skills

Each step below MUST be a separate conversation turn. Never combine steps.

Step 1: Understand the Need

Identify what domain and specific task the user needs help with.

Step 2: Search

Run find_skills.py with relevant keywords. Try multiple queries if the first doesn't yield good results. Even if the user names an exact skill, always search first to find the correct source/owner and verify it exists.

Step 3: Present Results and Ask

Show the user the found skills with:

Skill name
Source repository
Install count (popularity indicator)
skills.sh link for more details

Ask which skill(s) they want to install. End your response here — no tool calls.

Step 4: Confirm

When the user picks a skill, repeat back what you will install and ask for confirmation. This must be a text-only response with no tool calls. Wait for the user to reply.

Step 5: Install

Only after the user confirms in a separate message, run add_skill.py. Report the result.

Common Skill Categories

Category	Example Queries
Web Development	react, nextjs, vue, css, tailwind, html
Testing	testing, jest, playwright, cypress
DevOps	docker, kubernetes, ci-cd, terraform
Documentation	docs, readme, markdown, api-docs
Code Quality	lint, refactor, code-review, typescript
Design	ui, design, figma, accessibility
Data & ML	pandas, data-analysis, machine-learning
Productivity	git, automation, workflow

Tips for Effective Searches

Use specific domain keywords: "react performance" instead of just "fast"
Try alternative terms if first search yields few results: "testing" → "jest" → "playwright"
Popular skill sources include: vercel-labs/agent-skills, google-labs-code/stitch-skills
Check install counts — higher counts generally indicate more mature skills

When No Skills Are Found

Acknowledge that no matching skill exists yet
Offer to help the user directly with their task
Suggest the user could create a custom skill for their use case using skill-creator

skill-finder