Skip to main content

Tutorial: Discover skills

The Discover stage is where every skill journey begins. You ask Skillsmith to find skills that match an intent — by keyword, by trust tier, by quality score, or by what a similar developer is using — and it returns candidates ranked by relevance.

These tutorials show Skillsmith in Claude Code. For installation in your preferred runtime (Cursor, Continue, Copilot, Windsurf), see Getting Started.

What you will do

  • Search the registry by keyword
  • Filter results by trust tier and quality score
  • Get recommendations based on a project context
  • Recognize when Discover is the wrong stage and you should jump to Evaluate or Install instead

Tools that map to Discover

Two MCP tools and two CLI subcommands cover the Discover stage:

Surface Tool / command Purpose
MCP search Keyword search with optional category, trust-tier, and score filters
MCP skill_recommend Context-aware recommendations (e.g. "skills for a React project")
CLI skillsmith search Same as MCP search, formatted for the terminal
CLI skillsmith recommend Same as MCP skill_recommend, formatted for the terminal

Step 1 — Search by keyword

In Claude Code, type a natural-language prompt. The Skillsmith MCP server binds these phrases directly — you do not need to prefix them with "Use Skillsmith to..." any longer.

Try these prompts:

  • "Search for testing skills"
  • "Find git workflow skills"
  • "Show me DevOps skills above quality 80"

Claude Code will invoke the search tool and return a ranked list. Each entry shows the skill ID (author/name), a short description, the trust tier, and a quality score from 0-100.

Recognizing a successful search

You will see a tool-use indicator from Claude Code naming mcp__skillsmith__search. If you instead see a generic answer that looks like training-data summary, the MCP did not bind. Ask again with the explicit form "Use Skillsmith to search for ..." and check that the server is configured in ~/.claude/settings.json.

Step 2 — Filter by trust tier

Skillsmith ranks every skill into one of five trust tiers (see the Trust Tiers reference). For production work you usually want verified or curated; for exploration, community is fine.

Try these prompts:

  • "Search for verified testing skills"
  • "Find curated CI/CD skills"
  • "Show me only verified or curated skills tagged 'react'"

The MCP server accepts trust_tier values official, verified, curated, community, and unverified — exactly the labels documented on the Trust Tiers page. Skillsmith will pass your phrase through to that filter.

Step 3 — Filter by quality score

Quality scores combine SKILL.md structure, freshness, install velocity, and review signal. A score above 80 is generally safe; below 60 is experimental.

Try these prompts:

  • "Search for testing skills with a quality score above 80"
  • "Find DevOps skills, top 5 by quality"

Step 4 — Get recommendations for your project

Sometimes you know your stack but not what skills exist for it. Skillsmith's skill_recommend tool takes a free-text context and returns skills that other developers in similar contexts have installed.

Try these prompts:

  • "Recommend skills for my React project"
  • "What skills should a Node.js developer install?"
  • "Suggest skills for a team running CI on GitHub Actions"

Recommendations include co-install signals — if you ask for React skills, you will often see entries marked "frequently installed alongside" a popular React skill. That signal only appears once a pair has reached five or more co-installs.

Step 5 — Use the CLI for repeatable searches

If you script your skill discovery (for onboarding new team members, for example), the CLI gives you stable JSON output.

Run this in your terminal:

skillsmith search testing --trust verified --min-score 80 --limit 5
skillsmith recommend "react project with vitest"

The CLI's --json flag emits machine-readable output suitable for piping into jq or downstream tooling. See the CLI reference for the full flag set.

Common pitfalls

The MCP returns no results

First, broaden the query. The registry has roughly 14,000 skills; a five-word phrase often over-narrows. Drop adjectives, drop the min_score filter, drop the trust-tier filter. If you still get nothing, the registry may genuinely not have what you are looking for — at which point the Author tutorial is where you go next.

Claude answers from training data instead of calling the tool

This was the symptom that drove the Wave 1 tool-description rewrite. Bare prompts now bind, but if your runtime is older or running a stale cache, fall back to the explicit form: "Use Skillsmith to search for testing skills". If even that fails, run skillsmith setup in your terminal — that installs the bundled /skillsmith slash-command skill, which is the most reliable trigger on Claude Code.

Results look stale

The registry indexes four times a day (00, 06, 12, 18 UTC). New skills appear within hours of publish. If you see a skill version that is older than what the author's GitHub shows, the indexer has not yet run; wait for the next cycle.

Where to next

You have a list of candidates. Now you need to decide which one to install. That is the Evaluate stage — compare candidates side-by-side, read trust badges in detail, and diff versions if you are upgrading.

Reference: MCP search tool · CLI search command · Trust Tiers.