/skill-stocktake
Sweep the skill + agent corpus for content-quality issues a deterministic gate can't judge — the judgment half of skill-quality (the docs-sync v2 pattern). atl skills check proves the corpus is structurally sound; /skill-stocktake judges obedience (does a skill follow its own documented flow?) and redundancy (do two skills do the same job or contradict each other?).
It is both manually callable and auto-triggered. atl session-start signals "a stocktake is due" once asset-affecting commits (core/ teams/) have piled up since the last recorded stocktake, gated by a ~1-day runaway-guard — you run the skill then.
When to use it
- When
atlreports "a stocktake is due — run /skill-stocktake" at session start. - Any time you want to sweep the skill corpus on purpose (
--allforces the whole corpus).
How it works
Deterministic first
The skill runs atl skills check and fixes every FAIL (a missing frontmatter block, a team.json↔disk mismatch, a child with no summary) — mechanical, zero-false-positive. It never hand-judges what the CLI already proves.
Change-aware
A full-corpus sweep every time is wasteful, so it scopes to the skills/agents touched since the last stocktake (the cursor lives in ~/.atl/skill-stocktake-state.json). Nothing changed and no --all? There's nothing to sweep.
Semantic, grep-grounded, adversarial
Then two passes, each guarded against the ~40% multi-agent-audit hallucination rate:
- Obedience — read each in-scope skill's
SKILL.mdagainst itself: a step referencing a flag it never defines, a contradiction between description and body, a dangling "see step N", a promised output the flow never produces. - Redundancy — compare each in-scope skill against the rest: two skills claiming the same trigger, doing the same job, or giving contradictory instructions.
Every finding is grep-grounded (quoted verbatim, or dropped) and adversarially verified (challenged, and dropped unless it survives).
Proposes — never rewrites silently
Surviving findings are proposed via AskUserQuestion (fix a broken flow / merge two overlapping skills / reconcile a contradiction). A skill rewrite touches identity, so the human confirms it — this is the one place /skill-stocktake is stricter than /docs-audit, whose prose fixes are safe to auto-apply.
Records the stocktake
On completion the skill stamps the cursor (atl skills check --record-stocktake), which resets the runaway-guard so session-start won't re-signal for ~1 day.
The CLI / Skill split
/skill-stocktake is the LLM half of skill-quality. The deterministic half — frontmatter validity, team.json consistency, agent-KB children — is atl skills check, which also runs as a CI gate on every PR. The skill never re-derives what the CLI proves; it spends LLM effort only on obedience + redundancy, grep-grounded and adversarially verified.
Related
atl skills— the deterministic half this skill builds on./docs-audit— the same backstop shape for the docs site (its prose fixes are autonomous; this skill's are proposed).