Brand voice extraction

You produce a voice profile that’s specific enough to be falsifiable — observable patterns, not adjectives. The profile pairs with a reusable prompt library so future writing can be drafted in-voice and scored against the same rules.

“Warm, confident, knowledgeable” is not a voice profile. A voice profile is “second-person, present tense, mean sentence length 14 words ±5, contractions on, no em dashes, opens with a noun not the article.”

Inputs to gather first

1. 10–15 pieces of writing the brand is proud of — its high-water mark, not its full output. Strip everything else.
2. The author distribution — if 12 of 15 are by one person, the profile is that person’s voice, not the brand’s. Flag and ask: declare that person the canonical voice, or require a more diverse corpus.
3. Channels the voice will be deployed against — blog, email, social, landing pages. The profile may need to split into sub-voices if channels have meaningful differences.

If .lens/positioning-brief.md exists, read it for context on the brand’s overall positioning. Voice should be consistent with position.

The pipeline (four phases)

Phase 1 — Corpus curation

Confirm the 10–15 pieces are loaded. Reject thin corpora (<8 pieces) — the patterns won’t stabilise.

Phase 2 — Pattern extraction

For the full corpus, run this extraction prompt:

SYSTEM: You are a stylometrician analysing a brand's writing. You surface
measurable, observable patterns — not adjectives, not feelings. Every
pattern must be one a human editor could verify by counting or reading.

USER:
Corpus (JSON list of {title, body}): {CORPUS}

Return JSON in this exact shape:

{
  "sentence": {
    "mean_words": <number>,
    "variance": "<high|medium|low>",
    "longest_typical": <number>,
    "shortest_typical": <number>
  },
  "paragraph": {
    "mean_sentences": <number>,
    "typical_range": [<number>, <number>]
  },
  "punctuation": {
    "contractions": "<common|occasional|absent>",
    "em_dashes": "<common|occasional|absent>",
    "semicolons": "<common|occasional|absent>",
    "exclamation_marks": "<common|occasional|absent>",
    "questions": "<common|occasional|absent>"
  },
  "voice": {
    "person": "<first|second|third|mixed-which>",
    "tense": "<present|past|mixed-which>",
    "formality": "<formal|neutral|informal>",
    "jargon_level": "<none|domain-light|domain-heavy>"
  },
  "openings": {
    "typical_opener_pattern": "<plain English with 3 examples from corpus>",
    "openers_avoided": ["..."]
  },
  "rhetorical_devices": [
    {"device": "<name>", "frequency": "<common|occasional>", "example": "<verbatim>"}
  ],
  "transition_words": {
    "used": ["..."],
    "avoided": ["..."]
  },
  "what_this_brand_never_does": [
    "<observable patterns absent from the entire corpus>"
  ]
}

Rules:
- Counts must be plausible for corpus size given.
- "Avoided" lists are inferred only when the pattern is absent across
  the WHOLE corpus.
- Every example verbatim, copy-paste accurate.

Then count the numeric fields yourself. Sentence mean, paragraph mean and punctuation rates come from a short script or wc, not from the model’s estimate. Overwrite the numeric fields with the counted values; keep the model’s pattern fields. In the September 2026 retest the model’s paragraph-length estimate ran 27% high and the rubric inherited the error.

Phase 3 — Voice profile synthesis

Aggregate into the final profile. The what_this_brand_never_does list is the most underrated section — a voice is as much defined by what’s missing as what’s present.

Phase 4 — Prompt library build

For each of the team’s recurring output types (blog opener, social caption, email subject, sales follow-up, landing-page H1), build a drafting prompt template that constrains against the profile. Each template:

• Loads the voice profile as system context
• Includes the never-does list explicitly
• Names the channel constraints (length, format)
• Includes the per-template eval rubric (see below)
• Tells the drafting model not to add facts, places, numbers or product claims that aren’t in the brief

The voice rubric

Output a 12-check rubric for scoring future drafts. Each check is pass/fail or numeric:

Check	Threshold
Mean sentence length within ±20% of profile	Within range
Sentence variance matches profile	Match
Paragraph length within profile range	Match
Contractions present per profile rule	Match
Em dashes count at or below profile	Match
Semicolons count at or below profile	Match
Person matches profile	Match
Tense matches profile	Match
Opener follows profile pattern	Match
No words from “avoided” list	Match
At least one rhetorical device from profile	Present
Reading-level within profile bracket	Within range

Drafts scoring 10+/12 ship. 8–9/12 go back to the model with failing checks named. <8/12 regenerate from scratch.

Calibrate before shipping: score five corpus pieces. They should reach 10/12. If they don’t, soften numeric ranges to ±25%. If they still fail on pattern checks (opener shape, rhetorical device), those checks were fitted to a few examples — drop them or make them corpus-level (“in at least half of pieces”).

Ship the rubric as a runnable script (Python or the project’s runtime). The script takes ~30ms per draft, fast enough to gate a CI step.

Output

Three artefacts:

1. Voice profile — the JSON from Phase 3 plus a Markdown summary
2. Prompt library — prompts/ directory with one file per recurring output type
3. Voice rubric — voice-rubric.py runnable scorer + voice-rubric.json configuration

Save the voice profile to .lens/voice-profile.json. Downstream skills (eval-gated-drafting, social-content-factory, lifecycle-journey-builder) read this file.

Evals

Self-check before delivery:

• Profile is observable, not adjective-based — no “warm” or “confident” anywhere in the JSON
• Counts are measured — numeric fields come from a script, not the model’s estimate
• Avoided list is non-empty — strong voices have things they never do
• At least one rhetorical device captured with a verbatim example

Failure modes to watch

• Corpus dominated by one author — profile becomes that person, not the brand. Flag and decide.
• Model collapses to vibes — if the JSON contains “warm” or “confident,” regenerate with stricter system prompt. Claude tends to honour structured-output discipline; smaller models often need few-shot examples.
• Brand is a publisher — if athletes and guest writers produce most of the journal, the profile describes an editor’s taste, not a voice. Build the house profile from brand-written pieces; treat guest pieces as context.
• Two-voice brand — some brands have distinct blog-voice and product-voice. Build two profiles, route through the right one per channel.
• Voice drift after 6 months — re-run the extraction quarterly. Diff against previous profile. If the diff is significant, the brand has drifted and you decide whether to formalise or pull back.