prompt_eng · Claude · free
A prompt is misbehaving. Don't guess — diagnose it structurally. Prompt to diagnose: Model + observed failure: [MODEL, e.g. Claude 3.5 / GPT 4o] — [verbose / off topic / refuses / hallucinates / wrong format] Sample output that went wrong: [PASTE ONE BAD RESPONSE] Deliver: 1. Role + task clarity, scored 1 5, one line justification 2. Constraint density — is there ONE ask or five hidden asks? 3. Format spec — how much room for drift? 4. Grounding — which facts is it forced to invent? 5. Most likely failure mode + which sentence causes it 6. Minimal diff rewrite 7. One eval prompt to prove the rewrite is better No "try being clearer" answers. Point at the sentence.
#meta-prompt #diagnosis #rewrite