Brand voice

You capture a real person's or brand's writing voice from samples and turn it into a reusable profile the other skills write in. A captured voice is the most valuable context this collection holds. It overrides the generic "sound human" rules, because the goal stops being "avoid AI tells" and becomes "sound like this specific writer."

You're done when the profile holds only traits the samples show, each with a do and a tiny example, plus the verbatim anchor lines, and it's saved where the other skills read it.

Initial assessment

• Read .agents/blog-context.md if it exists. The Voice section is where the profile belongs; update it rather than starting a parallel file. For more than one author, use .agents/voice-<name>.md and note which posts use which.
• Get samples. Ask for three to five pieces the author actually wrote: posts, emails, threads, or a URL to their writing. More and longer is better. Without samples you can't capture a voice, so don't invent one. If the user has none, capture stated preferences by interview and label the profile aspirational until real samples exist.
• Check you have enough. Count the words you actually have. Under about 400, capture only the traits the samples clearly support, label the profile "low-confidence (thin samples)" in the Voice section, and tell the user which dimensions you couldn't observe (usually transitions and how they sustain a long middle). Never fabricate a trait to reach a count.

Method

1. Read for what repeats, not what it's about. Ignore the topics; study the moves. Read at least a few hundred words before you generalize, so you catch habits rather than one-off choices. Separate voice from belief: a trait must be a portable mechanic (sentence shape, person, punctuation, an opening move), not a claim only this author can make ("I've done this at four companies") or a topic-bound opinion. Capture the shape ("earns authority by citing specific reps"), never the specific claim, so the skills that write later never fabricate experience. Real stories, results, and stances you find in the samples are still valuable: record them in the "Experience and proof" section of .agents/blog-context.md, where writing skills may cite them as facts.
2. Profile the voice across concrete, imitable dimensions:
   • Rhythm: typical sentence length and how much it varies (short and punchy, long and winding, or a deliberate mix).
   • Formality and person: first or second person, contractions, slang, how close the writer stands to the reader.
   • Signature moves: rhetorical questions, fragments, asides, humor, analogies, direct address, strong openings. This is the fingerprint.
   • Vocabulary: plain or technical, the words they reach for, the words they avoid.
   • Stance: how much opinion, hedging, or certainty they show.
   • Structure: how they open, how they transition, how they land an ending.
   • Punctuation habits: dashes, parentheses, lists, emphasis. Note what they actually do, even when it breaks a generic rule.
3. Pull two or three short verbatim lines that are unmistakably theirs, to anchor the profile in real examples.

Output

If .agents/blog-context.md exists, replace its Voice section. If it doesn't, create it from blog-context's template and fill in only the Voice section, leaving the other sections as bracketed prompts for blog-context to complete later. For more than one author, write .agents/voice-<name>.md. The Voice section holds:

• As many copyable traits as the samples genuinely support (up to ten; fewer is fine for thin samples), each with a one-line do and a tiny example. For instance: "Opens with a blunt claim, not a windup. Do: 'Most cold brew advice is wrong.' Not: 'Cold brew has become increasingly popular.'"
• A short "keep these, they're intentional" list: quirks a generic pass would flag as tells but that belong to this voice (a habit of fragments, a signature dash, a rule-of-three). Name each one with the heading ai-tells.md uses for it ("Fragments posing as paragraphs," "Question, then a clipped answer," "Stacked threes," or "em dashes" for hard rule 1), and add the line "These win over ai-tells.md wherever they conflict." Matching names are what make humanize and article skip them instead of sanding them off. A voice can keep a habit, never a fabrication: the no-fabrication and chatbot-residue rules still apply.
• The two or three verbatim anchor lines.

Confirm the path, then tell the user that article, humanize, and the rest now write in this voice and that it overrides the generic anti-AI rules wherever they conflict.

Related skills

blog-context · humanize · house-style · article