Reverse-engineer a likely prompt from an output you want to reproduce

Structural signature: every section opens with a one-sentence plain-language summary before any detail, appearing in all 4 sections consistently, this pattern recurring across every section (not just once) suggests a deliberate instruction like 'start each section with a one-sentence summary before elaborating', not a coincidence. Likely incidental: the output uses the word 'crucially' twice, that's probably just the model's word choice for this generation, not evidence of a specific instruction to use that word, wouldn't build a prompt requirement around it. Candidate prompt: 'For each section, write a one-sentence plain-language summary first, then 2-3 sentences of supporting detail. Use concrete examples rather than abstract claims. Avoid absolute language, prefer measured statements.' Confidence: fairly confident about the summary-first structure, that's a strong, repeated signal. Less confident about the 'avoid absolute language' inference, that's based on noticing hedged phrasing in 2 of 4 sections, which could also just be topic-appropriate caution rather than an instructed style. Verification: run the candidate prompt on a similar but new topic, compare whether the new output also opens each section with a one-line summary and maintains similar hedging, if the structural pattern reproduces reliably, the reconstruction is likely close.