review-skills

name	review-skills
description	Page through the discovered-skills queue 25 at a time — keep the good ones into the verified registry, reject the rest so they never re-appear. Resumes where you left off.
version	1.0.0
author	veekunth217
tags	["review","triage","registry","discovery","maintainer","curation"]
platforms	["claude-code","cursor","codex"]

/review-skills — Triage the Discovery Queue

The weekly scraper drops every Claude Code skill it finds into registry/discovered.json, sorted by quality score. This skill walks you through that queue 25 at a time, so you can decide per skill: keep (→ verified registry) or reject (→ never shown again). Your decisions persist between sessions — next run picks up where you stopped.

RULE: Never modify registry/skills.json or registry/rejected.json without showing the user the exact change and getting GO.

---

Phase 1 — Where Are We

Run silently:

# Must be inside the claude-scaffold-skill repo (or have it as ~/.claude/skills/claude-scaffold-skill)
[ -f registry/discovered.json ] && echo "HERE=local" || \
  ([ -f ~/.claude/skills/claude-scaffold-skill/registry/discovered.json ] && echo "HERE=installed" || echo "HERE=missing")

# Status snapshot
python scripts/review-queue.py --status 2>/dev/null || \
  python ~/.claude/skills/claude-scaffold-skill/scripts/review-queue.py --status 2>/dev/null

If HERE=missing:

No discovery queue found. Run the scraper first:
  cd ~/.claude/skills/claude-scaffold-skill && make discover
(or trigger the weekly GitHub Action manually)

Then stop.

Otherwise, show the status block to the user:

Review queue
────────────
  In discovered.json:  [N]
  Already accepted:    [N]   (in skills.json)
  Already rejected:    [N]
  Still pending:       [N]   → [pages] pages of 25 left

Ready to review the next page? (yes / set page <n> / status / quit)

---

Phase 2 — Show a Page

Run python scripts/review-queue.py --page <n> --size 25 and present the output as a clean numbered list. For each candidate show:

[ 1] ⭐ 17,127  score 100   zarazhangrui/frontend-slides
     Create beautiful slides on the web using Claude's frontend skills
     → github.com/zarazhangrui/frontend-slides   ·   SKILL.md ↗
     flags: (none)

[ 2] ⭐  1,046  score 100   eugeniughelbur/obsidian-second-brain
     Claude Code skill for Obsidian — turn your vault into an AI-first second brain
     → github.com/eugeniughelbur/obsidian-second-brain   ·   SKILL.md ↗
     flags: (none)

...up to 25...

If a candidate has flags (low_stars, stale, thin_desc, non_canonical_topic, archived, thin_fork), surface them — they're hints to lean toward reject.

Then ask:

For this page (1–25), tell me your decisions. Examples:
  "keep 1 3 7, reject the rest"
  "keep 2, skip the others for now"
  "reject all"
  "open 4"  → I'll fetch and summarize that repo's SKILL.md so you can judge
  "next"    → move on without deciding (they stay in the queue)

Support open <n> — fetch https://github.com/<repo>/blob/HEAD/SKILL.md (raw), summarize what it does, whether it follows confirm-before-generate, scope, and any red flags in the install command. Then re-ask for decisions.

Default for anything not mentioned: leave it pending (no decision). Never auto-reject silently.

---

Phase 3 — Confirm the Batch

Before writing anything, show exactly what will change:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
PROPOSED CHANGES — page [n]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

KEEP → registry/skills.json (verified=true):
  + zarazhangrui/frontend-slides
  + dpconde/claude-android-skill

REJECT → registry/rejected.json (scraper skips forever):
  − some/low-effort-repo
  − another/archived-thing
  ...

LEFT PENDING (no change):
  [count] — will show up again next time

NOTHING is written until you type GO.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Type GO to apply, EDIT to change decisions, or CANCEL.

---

Phase 4 — Apply

On GO, run the persistence tool:

# Accepts (each one appended to skills.json as verified)
python scripts/review-queue.py --accept owner/repo --accept owner/repo2 ...

# Rejects (each one added to rejected.json)
python scripts/review-queue.py --reject owner/bad1,owner/bad2 ...

Then validate:

python scripts/validate-registry.py

If validation fails (e.g. a description over 150 chars, a tag with spaces), fix the offending entry in skills.json and re-validate. Common fixes:

• Trim description to ≤ 150 chars
• Lowercase + hyphenate tags, cap at 10
• Ensure added is YYYY-MM-DD

---

Phase 5 — Continue or Stop

✓ Page [n] done.

  Accepted this page:  [N]
  Rejected this page:  [N]
  Remaining pending:   [N]  ([pages] pages left)

Next?  (next page / set page <n> / status / commit / quit)

• next page → loop back to Phase 2 with page n+1
• commit → stage registry/skills.json + registry/rejected.json, show the diff, and on GO run:

  git add registry/skills.json registry/rejected.json
  git commit -m "registry: review pass — keep N, reject M"
  git push
  

  (If branch protection blocks the push, remind the user the weekly Action uses PAT_TOKEN; for manual pushes they push as themselves.)
• quit → remind them their decisions are saved in the files but not yet committed; they can resume /review-skills anytime.

---

Notes

• Rejected ≠ deleted. Rejected repos live in registry/rejected.json. The scraper skips anything listed there, so they won't clutter discovered.json on the next run. Un-reject by removing the slug (or by --accepting it later — that auto-removes it from rejected).
• Accepted skills get verified: true — that's the human-review stamp. The CI 404-check still runs on the PR.
• discovered.json is regenerated every scrape. Pending items persist because the scraper re-finds them; accepted/rejected ones drop out because the scraper skips repos in skills.json and rejected.json.
• This skill is the interactive front-end for scripts/review-queue.py — you can also drive that script directly (--page, --accept, --reject, --status, --all).