Rewrite your prompt into the native dialect of the model you're actually running (Claude Code / Codex) — by each vendor's OFFICIAL rules — and see which rule drove each edit. Self-updating. Zero deps.