bd rules audit|compact — “audit and compact Claude rules”: scan .claude/rules/ for contradictions and merge opportunities (Jaccard similarity, default threshold 0.6) and merge related rules into composites (--auto applies the audit’s suggestions, --dry-run previews, --group names them explicitly).
This is the odd one out on the page — a tracker with a command that operates on the harness’s guide layer, not on work at all — and it is the most interesting boundary crossing beads makes.
The problem it targets is real and well documented in topic-harness-engineering: a project’s rule set grows by accretion as agents and humans add instructions, and after a few months it contains rules that overlap, rules that contradict, and rules that are dead. Nobody rereads it, so it silently degrades, and it degrades the model’s behaviour with it. ce-compound-refresh is the only comparable capability in the wiki — keeping a corpus healthy over time (Keep / Update / Consolidate / Replace / Delete) — and it is a framework skill, an LLM asked to curate. Beads’ version is mechanical: a similarity metric, a threshold, a merge.
Why it lands in a store is not obvious and is probably not principled — it is a tool built by people who felt the pain in their own repo. But there is a defensible reading: rules are durable project state that outlives every session, which is this layer’s whole subject, and .claude/rules/ is exactly the kind of file beads-remember argues should not be maintained by hand.
Maps to no canonical SDLC stage.
See Also
- topic-harness-engineering — the guide layer this maintains, and why it decays.
- ce-compound-refresh — the LLM-driven counterpart for a solutions corpus.
- beads-setup — the command that writes into the same guide layer.
- beads-remember — the store-native alternative to hand-maintained instruction files.