find-skill
Scour the internet for proven, well-reviewed, high-quality agentic-CLI integrations (plugins, skills, hooks, agents, MCP servers, proxy layers, CLAUDE.md / AGENTS.md prompts) for a given topic. Survey
SKILL.md.claude/skills/find-skill/SKILL.md
GitHubfind-skill
Survey proven agentic-CLI integrations for a topic, score them, justify a pick, and offer a tailored hybrid synthesis.
/find-skill <topic> --scope <scope> [--n 5] [--inspire 1,3,4]
Scopes: agentic | plugin | skill | hook | agent | mcp |
proxy | project-md | user-md (multi-allowed: --scope plugin,hook).
Helper scripts live in scripts/. Canon source list in
references/canon-sources.md.
Output skeleton — render verbatim
Every run emits exactly these seven sections, in this order, with these banners. The user reads top-to-bottom; do not reorder, do not skip.
═══════════════════════════════════════════════════════════
/find-skill — <topic> | scope: <scope> | n=<N>
═══════════════════════════════════════════════════════════
▸ STEP 1 — Parse & confirm
▸ STEP 2 — Curated-list sweep
▸ STEP 3 — Community signal pass
▸ STEP 4 — Quality filter
▸ STEP 5 — Compare-and-contrast
▸ STEP 6 — Recommendation + justification
▸ STEP 7 — Hybridize? (interactive)
▸ STEP 1 — Parse & confirm
Echo the parsed inputs. If anything is ambiguous, ask one tight clarifier and stop until answered.
┌─ Inputs ──────────────────────────────────────────────────
│ Topic : <topic>
│ Scope : <scope>[, <scope>...]
│ N : <count> (default 5, max 10)
│ Inspire : <ids or none>
│ Workflow: <one line about user's current task — read from
│ conversation context, not invented>
└───────────────────────────────────────────────────────────
The "Workflow" line matters — it anchors STEP 6 ("choose this if…") and STEP 7 (hybrid intent). Pull it from what the user has been doing this session, don't fabricate.
▸ STEP 2 — Curated-list sweep
Run scripts/survey.sh <topic> <scope> for a one-shot fan-out across
awesome.ecosyste.ms, GitHub repo search, and Reddit. Cross-reference with
references/canon-sources.md to pick up pinned indexes the search may miss.
Render survivors compactly:
┌─ Curated lists & seed repos ──────────────────────────────
│ • <name> ★<stars> pushed:<date> <one-liner>
│ • …
└───────────────────────────────────────────────────────────
Hard cap: 8 lines. If more, demote the rest into STEP 3.
▸ STEP 3 — Community signal pass
Spawn parallel haiku agents (one Agent message, multiple tool blocks) when raw count > 8 candidates. One agent each:
| Agent | Task |
|---|---|
| reddit-haiku | Reddit threads for <topic> last 12mo, score≥5 |
| gh-haiku | GitHub issues/discussions: usage reports, fail modes |
| web-haiku | Blog posts / HN / dev.to last 12mo |
Each returns ≤200-word digest + URLs. Dedupe with scripts/dedupe.sh.
┌─ Community signal ────────────────────────────────────────
│ Reddit : <N> threads, top: "<title>" (r/<sub>, <score>↑)
│ GitHub : <N> reports, top: <repo>#<issue>
│ Web : <N> posts, top: <author>, <date>
└───────────────────────────────────────────────────────────
▸ STEP 4 — Quality filter
Apply the gate. Reject candidates failing ANY:
- Last commit > 12 months AND not in
references/canon-sources.md - < 5 stars AND no notable referrer
- Broken install / missing manifest / dead MCP server
- Solo-author + zero closed issues (likely abandoned)
Keep candidates with ≥1 of:
- Active maintenance (commits last 90d)
- Cited in ≥2 independent sources
- In Anthropic cookbook / official MCP registry / academic paper
- Notable maintainer / strong PR throughput
Render the rejection log:
┌─ Filter results ──────────────────────────────────────────
│ Kept : <N>
│ Rejected : <N> reasons:
│ ✗ <name> — stale (last commit YYYY-MM-DD)
│ ✗ <name> — broken install per <issue>
│ ✗ <name> — duplicate of <name>
└───────────────────────────────────────────────────────────
Rejection transparency matters — the user can override.
▸ STEP 5 — Compare-and-contrast
Render TWO tables. First is the spec sheet, second is the decision-helper.
5a. Spec sheet
┌─ Candidates ──────────────────────────────────────────────
│ ID │ Name │ Type │ ★ │ Pushed │ Signal
│ 1 │ <name> │ skill │ 1.2k │ 2026-04-10 │ canon + 3 reddit
│ 2 │ <name> │ plugin │ 480 │ 2026-03-22 │ HN front page
│ …
└───────────────────────────────────────────────────────────
5b. Tradeoffs
┌─ Tradeoffs ───────────────────────────────────────────────
│ ID │ Pros │ Cons │ Choose if…
│ 1 │ tight scope, fast, cached │ no MCP integration │ you want a single skill, low ceremony
│ 2 │ batteries-included, MCP+hook │ heavy; opinionated │ you want full plugin with hooks pre-wired
│ …
└───────────────────────────────────────────────────────────
For each candidate also drop a 3-line block:
#<id> <name>
approach : <one sentence on the architectural shape>
distinctive : <the load-bearing idea — what others lack>
weakness : <what it gets wrong / misses>
▸ STEP 6 — Recommendation + justification
This step is mandatory. Pick one, then justify both the pick AND every rejection in the shortlist.
┌─ ✅ Recommendation ───────────────────────────────────────
│ Pick: #<id> <name>
│
│ Why it fits <user workflow from STEP 1>:
│ • <reason 1 — concrete, ties to workflow>
│ • <reason 2>
│ • <reason 3 — ideally cites a specific feature>
└───────────────────────────────────────────────────────────
┌─ ❌ Why the others didn't fit ───────────────────────────
│ #<id> <name>: <one-line reason tied to workflow>
│ #<id> <name>: <one-line reason tied to workflow>
│ #<id> <name>: <one-line reason tied to workflow>
└───────────────────────────────────────────────────────────
Rejection reasons must be workflow-relative, not absolute. "Uses Python not bash" only matters if user's workflow needs bash. Surface that explicitly.
If no candidate is a clean fit, say so and skip straight to STEP 7.
▸ STEP 7 — Hybridize? (always ask)
Always ask. Even when the recommendation is strong, the hybrid path is where this skill earns its keep.
┌─ 🧬 Genetic mixing ───────────────────────────────────────
│ Want a custom hybrid that genetically mixes 2-4 of the
│ above into something tailored to <user workflow>?
│
│ Here is what a hybrid could look like for YOUR case:
│
│ Parents : #<id> + #<id> [+ #<id>]
│ New intent : <one sentence — what this hybrid does that
│ no parent does alone>
│
│ Inherited primitives:
│ • from #<a>: <load-bearing piece — frontmatter/triggers/
│ step structure/hook wiring/etc.>
│ • from #<b>: <complementary piece>
│ • from #<c>: <polish piece>
│
│ New genes (not in any parent):
│ • <thing genuinely added — e.g. cache layer, gating,
│ project-specific trigger, output format>
│
│ Why this combo > any single parent:
│ <one or two sentences on the additive value>
│
│ Reply with one of:
│ 1. "ship it" — generate the hybrid file now
│ 2. "tweak: <notes>" — adjust parents, intent, or genes
│ 3. "skip" — just take the recommendation
└───────────────────────────────────────────────────────────
If user says "ship it", invoke STEP 7-build.
STEP 7-build (on "ship it")
Extract primitives — for each parent, list its load-bearing pieces (frontmatter shape, trigger phrases, step structure, hook wiring, prompt scaffolds, output format). Use
scripts/extract-skill.sh <url>per parent to keep it cheap.Conflict map — render points of disagreement:
┌─ Conflicts ───────────────────────────────────────── │ Trigger style : #1 keyword-list vs #2 regex → keep #1 │ Hook event : #1 PreToolUse vs #3 Stop → keep #3 │ Output format : #1 markdown table vs #2 jsonl → keep #1 (user reads) └─────────────────────────────────────────────────────Selection rationale — one line per primitive on why the winner won.
New synthesis — write a draft of the hybrid (skill / hook / agent) inline in the same file format the user would commit. Mark borrowed lines with
# from #<id>comments where non-obvious.Delta callout — bullet what is genuinely new. If nothing is new, say so and recommend just using the strongest single parent.
Hybrid count guidance:
| N parents | Pattern |
|---|---|
| 2 | Frame from one, fill from the other. Cleanest. |
| 3 | Frame + scaffold + polish split. Sweet spot. |
| 4-5 | Risk of Frankenstein. Force-rank, drop weakest. |
| 6+ | Don't. Cluster first, hybridize cluster reps. |
Then ask before writing to disk:
Save hybrid to ~/.claude/skills/<name>/SKILL.md ? [y/N/path]
Helper scripts (in scripts/)
| Script | Purpose |
|---|---|
survey.sh |
One-shot fan-out: awesome+gh+reddit, deduped. Start here. |
awesome.sh |
awesome.ecosyste.ms list search, TSV out |
gh-search.sh |
Recent active GH repos by topic+scope, pipe out |
gh-quality.sh |
Per-repo signal: stars, push, issues, archived |
reddit.sh |
Top reddit threads, no-auth, pipe out |
extract-skill.sh |
Cheap structural digest of a SKILL.md / agent.md |
dedupe.sh |
Dedupe pipe-rows by name, keep highest-stars |
All scripts emit pipe-delimited rows (or TSV). Token-shaped — feed to
grep to filter further; they are already compact, so do not run them
through another summarizing reader.
Output discipline
- Render the seven banners verbatim. Skip none.
- Lead each step with the table/box, prose after.
- Cite every claim with URL or candidate ID.
- Recommend zero candidates if none pass STEP 4 — do not pad.
- STEP 6 justification is non-negotiable. STEP 7 hybrid prompt is non-negotiable.
- Web access fails → degrade to
references/canon-sources.mdand say so.
Anti-patterns
- ✗ Recommending a candidate without fetching its README.
- ✗ Padding to N=5 with low-quality fillers. n=3 strong > n=5 mid.
- ✗ Hybridizing without showing the conflict map.
- ✗ Justifying the pick without justifying the rejections (asymmetric reasoning hides bias).
- ✗ Using Claude-in-Chrome for this — WebFetch / firecrawl_search / the helper scripts are faster.
- ✗ Writing the hybrid file without the "Save to ...?" confirmation.