knowledge-auditlisted

# /knowledge-audit — curate the autonomous knowledge lane The `Stop` hook (`.claude/hooks/auto-learn.sh`) writes candidate knowledge to `.claude/knowledge/auto/` **without human approval**. That speed has a cost: machine-written inferences can be wrong, stale, or duplicated. This skill is the **quality gate that runs after the fact** — it verifies each candidate against the live codebase and either promotes it to the curated KB or removes it. Without periodic auditing the auto lane rots and poisons every agent that reads it. > **Authority**: [`AGENTS.md`](../../../AGENTS.md) is binding. **This audit is normally run > autonomously** by the `Stop` hook (`.claude/hooks/auto-learn.sh` → `auto-audit-prompt.md`) > when the auto lane crosses `AUTO_AUDIT_THRESHOLD`. That unattended pass **defaults to observe-only** > (prune + mark `PROMOTE-CANDIDATE`; no curated writes unless its operating mode enables promotion). > This skill is the **on-demand manual variant that actively promotes**: it prunes **and** promotes > without human approval — the approval gate is replaced > by a **code-verification gate**: only candidates whose referenced files/symbols are proven to exist > (via grep/Glob) may be promoted to the curated KB. Edits to `AGENTS.md`/skills/agents remain > out of scope. Safety = code proof + git history + easy rollback (`rm auto/*.md` / `git revert`), > not human sign-off. > **Reference loading (progressive disclosure):** > Read [principles](reference/principles.md) first,