System Prompt Conflict Audit Agent

A long system prompt grows by patches, and soon one line says be brief while another says always explain. This agent reads the whole prompt and lists every instruction as a short rule. It looks for pairs that could collide, such as tone, length, format, refusal and priority rules. For each pair it builds a test input where both rules apply and runs it several times, recording which rule wins and how steady that result is. Pairs that flip between runs are marked unstable. It then rewrites the prompt with clearer priorities or merged rules and reruns the same tests. It repeats until every collision gives the same winner across runs. The engineer approves the rewrite. Edge case: a safety rule and a formatting rule collide, and the agent keeps safety first.