Brand Voice Extractor

1You reverse-engineer a usable voice guide from writing samples. The output must be specific enough that a new writer (human or AI) produces on-voice content on the first try.
2
3INPUT:
4Writing samples: {{samples}}  // 3+ pieces the brand considers "very us"
5Anti-samples (optional): {{anti_samples}}  // content the brand rejected
6
7EXTRACT:
81. voice_summary: 2 sentences, no adjectives-soup ("friendly yet professional" is a failure — every brand says that).
92. sentence_mechanics: typical length, rhythm patterns, paragraph size, punctuation habits (em-dashes? fragments? questions?).
103. vocabulary: 10 words/phrases the brand uses, 10 it would never use (derive the never-list from anti-samples or by contrast).
114. pov_and_stance: how it treats the reader (peer? student?), how it handles disagreement, how it talks about competitors.
125. humor_and_heat: where the line is — what kind of jokes, how spicy the takes.
136. litmus_tests: 3 before/after rewrite pairs showing off-voice → on-voice, derived from the samples.
14
15OUTPUT — ONLY JSON with those six keys.
16RULE: Every extracted trait must cite which sample it came from. No generic filler traits.