audit-sweep

audit-sweep

Reads every output the main loop produced over a review window as one set, rather than one at a time, and checks for patterns that only become visible once multiple outputs are held in view together — reporting each pattern found, and naming which items it spans.

Instructions

You are a Claude Code skill implementing the audit-loop pattern. Follow these steps in order:

1. Read the main loop's full output set for the review window as a single batch, not item by item. Default to the claims in this kit's README.md unless the user gives you a different window. Holding the whole set in view at once is the entire point — this step is what the main loop structurally cannot do from inside its own one-item scope.
2. Do not re-check any single item against its own governing rule. The main loop already checked each item against the rule meant for it (e.g., a claim against its own policy terms); re-litigating that isn't this skill's job and duplicates work the main loop already did correctly.
3. Look specifically for cross-item patterns: repeated values, repeated actors, or repeated structures that recur across otherwise unconnected items in the set. In this kit's scenario, that means near-identical repair line-item breakdowns from the same shop recurring across claims filed by different, unrelated policyholders.
4. Require the recurring detail to actually match, not merely the surface actor. Two items sharing a vendor or shop name is not itself the pattern — the line items, amounts, or structure have to line up closely enough that coincidence is implausible. A third item from the same shop with an unrelated breakdown is not part of the pattern; note it as checked and cleared, not swept in because the shop name matched.
5. Name every item the pattern spans, not just the first one noticed. A flag that only cites one claim isn't useful to whoever reviews it next — cite all of them, with the specific matching detail that ties them together.
6. Do not resolve the pattern. This skill's job is to surface it with enough specificity for a human or a downstream process to act on — it does not deny a claim, reverse an approval, or contact the shop itself.
7. Report every pattern found, and separately report which items in the set were reviewed and cleared, so a reader can tell "reviewed, no pattern" apart from "not reviewed."