Skill finder
Ranks your installed Claude Code skills/agents for a task and routes to the best one — and learns from your ratings.
npx -y skills add RamanaNani/agent-router --skill skill-finderAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Browse and discover Claude Code skills, agents, and plugins from the marketplace — ranked by rating — that are NOT yet installed. Use when the user is new and has few/no specialists installed, asks "what plugins/skills are available for X", "find me a skill for Y", "what should I install", "browse the marketplace", or when agent-router finds no installed tool that fits. Returns a ranked shortlist with real ratings, source links, and the exact install command.
SKILL.md
4.1 KB, as published. Nobody here has run it
Skill Finder — marketplace discovery
Where agent-router routes among tools you ALREADY have, skill-finder discovers
tools you DON'T have yet. It searches the marketplace, ranks candidates by rating,
and hands back the exact install command. Built for a new user whose ~/.claude/
is empty or thin.
When to use
- A new user wants to know what's worth installing for their kind of work.
- The user asks "find/browse skills for X", "what plugin does Y", "what should I install".
agent-routerfound nothing installed scoring ≥ 25 → fall through to here.
Procedure
1. Capture the need
One sentence: the domain + task (code review, testing, security, infra, data, research, design…). Note any language/framework constraint.
2. Don't recommend what they already have
Quickly scan installed tools (the session list, ~/.claude/, ./.claude/, and
installed plugins) so the shortlist is all NEW installs. If a good specialist is
already installed, say so and route to it instead (defer to agent-router).
3. Gather candidates — real ones only
Merge two sources:
- Curated catalog:
data/marketplace.json(verified entries + ratings). - Live search (preferred, for freshness):
- GitHub: repositories matching
claude code <domain> plugin|skill|agent, sorted by stars (use the GitHub search tool or a web search). - Web:
<domain> claude code skill OR plugin OR subagent. For each candidate capture: name, what it does, a rating signal (GitHub stars or curated score), the source URL, and how to install it.
- GitHub: repositories matching
NEVER invent a repo, package, or star count. List only candidates that appear in the catalog or in a real search result, and include the URL so the user can verify.
4. Rank (0-100)
- Relevance (0-50): match to the task/domain.
- Rating (0-35): GitHub stars (log-scaled) or curated
rating. Unknown → 10. - Maintenance (0-15): recent commits / not archived. Unknown → 7.
Drop anything below relevance 15 — discovery should stay on-topic.
5. Present the shortlist
Top 3-5 as a table: rank, name, what it does, rating (e.g. ★ 1.2k or score 80),
source link. Then, for the top pick, show the exact install command:
/plugin marketplace add <owner>/<repo>
/plugin install <plugin>@<marketplace>
or npx <package> when the project ships an npm installer.
6. Offer to install
Installing plugins is done by the USER via /plugin … — you cannot run those
slash commands for them. Lay out the 1-2 commands and offer to walk through it.
If the pick is an npx installer, you may run it via Bash once the user confirms.
7. Log the discovery (internal dogfooding)
Append ONE JSONL line so gaps in the installed toolset are tracked over time:
mkdir -p ~/.claude/agent-router/logs
echo '{"ts":"'"$(date -u +%FT%TZ)"'","skill":"skill-finder","need":"<one-line need>","domain":"<domain>","recommended":"<name|->","source":"<url|->","rating":"<stars|score|->","installed":false,"feedback":""}' >> ~/.claude/agent-router/logs/decisions.jsonl
Set installed to true once the user installs it. These rows reveal which
capabilities the user repeatedly needs but doesn't have — the roadmap.
Rules
- Honesty over coverage: a short list of real, linked tools beats a long invented one.
- Always show the source URL and the install command — discovery is useless without them.
- State the rating's basis (GitHub stars vs. curated score) so it isn't mistaken for gospel.
- If live search is unavailable and the catalog is thin, say so and show what you have.
- After install, hand off to
agent-routerto actually route the task.