connorads Skill Freshness
The lint owns phrasing; this sweep owns truth. A claim rots in the world, not the text - no grep can notice that a linter count grew or a flag was renamed. The sweep re-derives each checkable claim from its live source and proposes fixes for whatever drifted.
Scope is the authored catalogue only: ~/skills (public),
~/.config/skills/personal, and ~/.config/skills/private (a standalone
repo - commit its fixes there, not via dotfiles). Vendored skills are
upstream's problem - refresh them via update-vendored-skills, never rewrite
them here.
1 - Inventory targets
Anchor on the markers the lint deliberately tolerates:
rg -ni 'current as of|as of|last verified|verified against' \
~/skills ~/.config/skills/personal ~/.config/skills/private --glob '*.md'
(Case-insensitive and no forced year, mirroring writing-skills' banner regex - "As of July 2026" and dateless banners must both surface.)
Anchors are the floor, not the list. For each skill under review, read SKILL.md (and any reference it leans on) and collect every checkable claim: commands and flags, counts ("142 linters"), version literals, API fields, URLs, and prices. Skip judgement content - stances and trade-offs have no live source to check against.
2 - Verify against live sources
Work skill by skill; for a whole-catalogue sweep, batch ~5 skills per subagent so claim tables never share the orchestrator's context. Per claim:
- Tool behaviour: run the command (
--help,<tool> builtins, a dry run). - Versions/counts: query the registry or installed binary, not memory.
- URLs: fetch; a redirect to a new canonical home counts as drift.
- Unverifiable here (needs auth, another OS): record "unchecked", never assume pass.
Verdict per claim: confirmed / drifted (with the observed value) / unchecked.
3 - Re-baseline against the model
Models improve underneath skills. For swept skills that bundle evals/, run
the eval prompts without the skill (fresh context) and note what the model
now does unaided - that content is retirement material, per writing-skills.
This tier is expensive: sample it (skills touched recently, or the oldest),
don't force it on every sweep.
4 - Propose, don't rewrite
Output one table per skill: claim, source checked, verdict, proposed fix. Then apply the curation rule:
- Trivial verified fact (a count, a renamed flag, a moved URL - spot-checked live): fix directly, one commit per skill.
- Anything judgement-shaped (retiring content, restructuring, re-tiering): propose the diff and stop. Tier and scope are curation calls.
Prefer fixes that end the claim's rot class over refreshing its value: a
live-query pointer (hk builtins) beats an updated count, and a dated as-of
caveat beats an undated one.