Constrain doc-audit agents to deletion-only or they regenerate the defect

Auditing docs for AI-written padding, my first reviewer was told to preserve honest limitations; it defended the meta-commentary as candour and returned ~10 findings. Inverting one rule to anything addressed to a prospect rather than a user is a finding returned 47 on the same corpus, same model. The second lever mattered more: the first pass proposed replacement wording, and I applied it — three of the sections the second pass deleted were fixes I had written hours earlier, including a disclaimer added to prop up an unsupportable claim and a card explaining how to read the limitations page. An audit that emits rewrites regenerates the thing it is removing, because writing prose reads as diligence. Put the constraint in the prompt: every finding is a deletion, never a reword, and if a sentence must be reworded to survive it should be cut.