social-voice

Turn real writing samples into voice rules concrete enough that any future draft can be mechanically checked against them.

Context

Read social-context.md at the project root (also check .agents/social-context.md) — you will be updating its ## Voice section, and its Positioning, Audience, and Never sections tell you which register matters. If the file doesn't exist, offer to run the social-context skill first, but don't block: ask two inline questions (who is the audience, which platform matters most) and proceed; you'll create the file with only a ## Voice section at the end.

Workflow

1. Gather samples immediately. Ask the user to paste 3–10 pieces of their real writing, or point you at files to read. Best sources in order: published posts on their primary platform, emails they wrote to humans they like, blog posts. Reject samples that were AI-generated or heavily edited by someone else — ask "did you write these yourself, start to finish?" If you get fewer than 3, proceed but flag lower confidence.
2. Separate signal from context. Note each sample's medium — a LinkedIn post and a customer email have different formality baselines. Analyze the invariants: what stays the same across mediums is the voice; what changes is the format.
3. Measure the mechanics — actually count, don't vibe:
   • Sentence length: median words per sentence, and the range. Any one-word sentences?
   • Paragraph shape: one-sentence paragraphs? Walls of text? Where do line breaks fall?
   • Punctuation: em-dashes, semicolons, ellipses, exclamation marks, parentheses — count per 100 words.
   • Emoji: which ones, how often, positioned where (inline, end of line, never)?
   • Case: any lowercase-on-purpose? ALL CAPS for emphasis? Bold?
4. Extract the vocabulary fingerprint:
   • 5–10 words or phrases they reach for repeatedly.
   • Words they conspicuously avoid (corporate verbs? jargon? profanity?).
   • Whether they say "I", "we", or neither.
5. Study openers and closers separately — these carry the most identity. How do first lines start (a claim? a scene? a number? never a question?)? How do pieces end (a question to the reader, a flat statement, a sign-off phrase, nothing)?
6. Locate the humor and heat register: do they joke, and how (dry, self-deprecating, absurdist, never)? Do they take positions ("X is wrong") or hedge ("it depends")? Note the strongest opinion in the samples verbatim as a calibration example.
7. Draft the rules. Write 8–15 rules in must/never form, each one checkable by a machine or a stranger.
   • Good: "never opens with a question", "one-sentence paragraphs, max 2 sentences", "no exclamation marks", "em-dash once per post, max", "signs off with just the first name".
   • Bad: "conversational", "authentic", "punchy". Include 2–3 short verbatim quotes from the samples as calibration anchors.
8. Verify by imitation. Take one of the user's samples, reduce it to a 1–2 line content summary, then rewrite it from that summary using only your drafted rules — without looking back at the original. Show the rewrite next to the original and ask: "Does the rewrite sound like you? What's off?" Every "what's off" answer is a missing rule — add it, and if the user names two or more things off, run the imitation test once more on a different sample.
9. Before writing, confirm the draft clears every row of the Quality bar — send yourself back to the step that fills any gap. Then write the rules into the ## Voice section of social-context.md. Preserve anything already there that you didn't derive this session (slider values, admire/avoid accounts from the social-context interview) — append and reconcile, don't replace wholesale. If a new rule contradicts an old line, show both and ask which wins.