what-skill

what-skill

Match a software-development task to the best skill from five packs (Pocock, Osmani, pstack, gstack, superpowers), then give the exact command to install it. Anything outside these packs belongs to the find-skills skill (npx skills find <query>), not this one.

$SKILL_DIR is the directory containing this SKILL.md. It ships two references:

• $SKILL_DIR/references/catalog.md: every publicly installable skill from the five packs, grouped by category, with pack, a one-line description and the install command.
• $SKILL_DIR/references/packs.md: a cited comparison of the big packs (Pocock, Osmani, pstack, gstack, superpowers): what each believes in, verbatim install commands, what the comparison videos found, and an overlap map (section 6).

Workflow

1. Restate the user's job in one line and map it to one catalog category (for example Planning & requirements, Debugging, Testing & verification). If the job is not software development, stop and use the find-skills skill instead.

2. Check the Pack picker below and the overlap map (section 6) in $SKILL_DIR/references/packs.md.

3. Search the catalog for the job's keywords (shell grep or your agent's search tool), starting with the mapped ## <Category> section. If no catalog row fits, say so and hand off to the find-skills skill; do not recommend skills from other sources.

   grep -i -E 'tdd|test-driven' "$SKILL_DIR/references/catalog.md"
   

4. Check what is already installed: run npx skills list (project) and npx skills ls -g (global), and look in .claude/skills/, .agents/skills/ and ~/.claude/skills/. If a suitable skill is already installed, say so and do not suggest installing it again.

5. Answer with one primary pick and at most two alternatives. For each, give:

   • the skill name and its pack,
   • one line on why it fits (cite the video or README from packs.md when the reason comes from there),
   • the install command from its catalog row. Add -g to install globally. If the pack has a Claude Code marketplace, also give its /plugin commands from packs.md.

6. Do not run install commands unless the user explicitly asks you to. When they do, run the exact command you showed.

Pack picker

Situation	Pick	Source
Big, vague feature with time to research first	Pocock wayfinder → to-spec → to-tickets	V2 (The Gray Cat)
Everyday feature work; build shared vocabulary as you go	Pocock grill-with-docs, tdd, diagnosing-bugs	V2, V3
Predictable full lifecycle with review and ship gates	Osmani /spec /plan /build /test /review /ship	V2
Hard bug, perf work, deep single-agent rigor (Cursor or Claude Code)	pstack /poteto-mode	V5 (Ray Fernando), pstack README
Role-based "virtual team" sprint (CEO/eng/design/QA/ship)	gstack	gstack README
Fast plan-then-subagent execution	superpowers	V3

Video IDs (V1–V5) and README sources are listed in packs.md section 1.

Rules

• Pick one pack per concern. Overlapping packs (for example Pocock tdd and Osmani test-driven-development) fight over the same triggers. Pocock's README says "Pick one".
• Do not install every skill. V1 calls running every skill on every project "overkill".
• Whatever the pack, the user still has to check the app by hand (V2, V3).
• The per-job verdict of V1 is unverified because it has no transcript. Never quote it as fact.
• gstack and superpowers install as whole packs (gstack via git clone + ./setup, superpowers via /plugin install). The catalog shows that pack command on each of their rows.

Refreshing the catalog

python3 "$SKILL_DIR/scripts/build-catalog.py" --root <dir with your clones>

--root is the directory that holds your skill repo clones (it can also come from the SKILLS_ROOT environment variable). The script rescans it for SKILL.md files and rewrites references/catalog.md. It keeps only skills it can trace to a public GitHub repo. It needs network access to confirm each source is public, and it exits with code 2 rather than write a partial catalog. It is only useful on a machine that has the clones. If the date on the catalog's Generated … on line is more than 30 days old and the root exists, offer to rerun it. It keeps only the five packs listed in ALLOWED_PACKS.