Back to Lobehub

Skill Freshness

.agents/skills/deep-review/references/dimensions/skill-freshness.md

2.2.101.9 KB
Original Source

Skill Freshness

Agent skills are load-bearing documentation: agents follow them literally. A diff that changes behavior a skill describes silently poisons every future agent session. This dimension keeps the skill library in sync with reality and spots knowledge worth capturing. Findings are advisories — they skip the verify pass.

Quick checklist

  • Does this diff invalidate any skill? Renamed/moved files, changed commands, altered workflows, removed exports that a skill references by path or name
  • Does this diff invalidate any statement in AGENTS.md / CLAUDE.md / README setup docs?
  • Is this diff the kind of foundational work later tasks will build on (new subsystem, new convention, new tool) with no skill capturing it?
  • Did this change fix a mistake class that keeps recurring? Root-cause knowledge belongs in a skill (or an update to the relevant dimension file of this review skill)

How to check

  1. Extract the diff's "renames and removals": file paths, exported names, npm scripts, CLI commands, env vars.
  2. rg each old name across .agents/skills/ and root-level agent docs (AGENTS.md, CLAUDE.md) — a hit on a removed/renamed thing is a stale-skill finding.
  3. For new subsystems: check whether any existing skill's scope covers it; if none does and the subsystem has non-obvious workflow (setup steps, conventions, gotchas), propose a skill with a one-line scope.

Violations (advisories)

  • A skill or agent doc that now states something false, with the exact file and line quoted.
  • A clearly skill-worthy gap: repeated-mistake class or foundational feature with no home. Propose the skill name and 2-3 bullet contents — concrete enough to act on, not "consider documenting this".

Not violations

  • Skills that are merely terse or could be "improved" without being wrong — freshness, not quality review.
  • Speculative skills for one-off work unlikely to recur.