ucsb-nlp-chang/skill-usage/find-skills-semantic

Finding Skills

Searches a local index of agent skills using semantic search to find the most relevant ones for a given task. Skills are pre-downloaded — no installation needed.

When to Use

• Starting a new task that may benefit from specialized skills
• Looking for best practices, patterns, or workflows for a specific domain
• Wanting to find tools or templates for a task (testing, deployment, design, etc.)

Search API

Use curl -s to query the semantic search endpoint.

Semantic search

Best for conceptual queries where you describe what you need in natural language.

curl -s "http://localhost:8742/semantic?q=QUERY&top_k=10"

Example: q=help me build and deploy containerized applications

Response format

The search endpoint returns a JSON array. Each result contains:

• name: skill name (may be duplicated across authors)
• description: what the skill does
• skill_md_snippet: first 100 words of the skill's documentation
• skill_id: unique identifier (author--name format), use with the detail endpoint
• github_stars: popularity of the source repository
• score: relevance score, 0-1 where higher = better match.

Skill detail

Fetches full metadata and complete SKILL.md content for a skill. Pass the skill_id from search results.

curl -s "http://localhost:8742/detail/SKILL_ID"

Workflow

Step 1: Analyze the task

Break the task into concrete sub-tasks. For example, "build a REST API with auth and tests" becomes:

• Design API endpoints and routing
• Implement authentication
• Write tests

Step 2: Search for each sub-task

For each sub-task, run semantic search queries to find skills.

# Sub-task: implement authentication
curl -s "http://localhost:8742/semantic?q=implement+authentication+with+JWT+tokens&top_k=10"
 
# Sub-task: write tests
curl -s "http://localhost:8742/semantic?q=writing+unit+tests+for+web+applications&top_k=10"

Refine queries if initial results are too broad or miss the mark. Try different natural language descriptions.

Step 3: Review and select

From the search results, select 10 skills total across all sub-tasks. Prioritize:

• High relevance to the target task
• Higher github_stars when multiple skills cover the same topic
• Skills with informative skill_md_snippet content

If needed, fetch full details of a skill to confirm relevance:

curl -s "http://localhost:8742/detail/SKILL_ID"

Step 4: Record results

Record the selected skills as a structured list:

## Found Skills
 
- **[skill-name]** (skill_id: [skill_id]) - [one-line summary from description]
- **[skill-name]** (skill_id: [skill_id]) - [one-line summary from description]
- ...