Using skills
Skill jasonm4130/claude-skills/plugins/superpowers-core/skills/using-skills
Personal Claude Code plugin marketplace: 9 plugins for orchestration guards (workflow-model-guard), tiered deep research (deep-dive), subagent-driven development, ADRs, session retros, handoffs, and adversarial review.
npx -y skills add jasonm4130/claude-skills --skill using-skillsAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 2 stars2 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
Use when starting any conversation — establishes the skill-selection and currency-verification rules that govern every turn (match-and-proportion, specificity-wins, user-instructions-suppress-skills, verify-before-answering-from-stale-memory). Delivered automatically via the SessionStart hook. Do NOT use for authoring new skill descriptions (see writing-skills) and do NOT use as a substitute for actually running a search or deep-dive — this skill states the decision rules, not the research mechanics.
SKILL.md
2.5 KB, as published. Nobody here has run it
using-skills
Part A — Skill selection (match-and-proportion):
- Invoke a skill only when its description is a closer, more specific match to the task than acting directly — and the work is non-trivial (multi-step, irreversible, or touches product source). Trivial/one-shot → just do it, no ceremony.
- Specificity wins: when two skills match, prefer the narrower one. Owned workflow skills are more specific to this work than any generic one → they win. (Structurally dissolves collisions — no future deny-line needed.)
- User instructions suppress skills: a concrete instruction in the current turn overrides any skill firing for the same concern. User > skills > defaults.
- When invoking, announce "Using [skill] to [purpose]" and follow it.
- Authoring rule: every skill description carries a negative scope ("do NOT use for…") to keep selection unambiguous.
Part B — Currency & verification:
- If an answer turns on something that changes over time (versions, prices, releases, "current/latest/now", anything plausibly past the training cutoff) → don't answer from memory; verify first.
- Match the tool to the need: one load-bearing fact → a single web search; a multi-angle/comparative/"state of X" question →
deep-dive. Never fan out a deep-dive on what one search settles. - Stale-fact ≠ unclear-intent: a factual gap you can close → verify it; an intent you'd only guess at → ask one question. State verified-vs-remembered when load-bearing.
Injection guardrail (keeps the kernel from re-becoming the superpowers wall). A behaviour earns a line in the SessionStart injection only if both: (a) it must be present before the model acts, and (b) it isn't already reliably governed elsewhere (global CLAUDE.md, or self-evident). Skill-selection and currency-verification pass; the Karpathy-4 / simplicity / verification-before-complete rules fail (b) — they live in global CLAUDE.md and must not be duplicated here.