Salience Voice

Voice

The failure mode this module prevents: a profile and a feed that are technically excellent and audibly not the person. At executive level that is expensive — the audience is specifically evaluating judgment, and outsourced-sounding writing reads as outsourced thinking.

Voice is how things are said. Positioning is what. Keep them separate; conflating them produces copy that sounds right and says nothing.

Two operations

Do not capture when a voice profile already exists and the user has not asked to rebuild it. Do not enforce personal voice on copy the user has said should read neutral — a company-page post or a formal board communication is a legitimate exception, not a voice failure.

Enforcement runs on every piece of writing this system produces, whether or not the user asks.

Capture

Samples

Enough genuine writing to see a pattern. Three or four substantial pieces works; 10-20 short posts works better. A pattern in one sample is a coincidence.

Two source families, and they capture different things. Use both where possible.

Raw sources give the truer voice. Slack messages, unedited emails to peers, call or podcast transcripts, anything written while annoyed or convinced. The less edited, the more of the actual person survives.

Published sources give the right register. The user's own past LinkedIn posts — ideally Shares.csv from a data export, which is free, large, and zero-effort. These have been shaped for an audience, so they carry less raw voice, but they show how the person writes in the destination.

Prefer raw sources for the underlying voice and published posts for calibration. Say which is which in the profile — a voice built only from polished writing will be flatter than the person, and one built only from Slack will be too loose for a profile.

Poor sources: anything ghostwritten · press releases · copy already AI-assisted · anything committee-edited · a résumé

Minimum gate: roughly 500 words. Below that, stop and ask for more rather than extracting a profile that is really a guess:

"That's too little to find a reliable pattern. Two or three more — old emails, Slack, a transcript — and the messier the better."

Exclusion check

Before extracting, identify what in the samples is not the person:

• Platform conventions mistaken for voice — LinkedIn's one-line paragraphs, hashtag habits, the short-line cadence the format encourages
• Quoted or borrowed phrasing from someone else
• Unusually formal registers — a legal disclaimer, a press release, board minutes
• Typos and autocorrect artifacts

This step matters most with LinkedIn samples specifically. Extract voice from a feed without it and you encode LinkedIn's house style as if it were the person's.

Ask directly: "Send me three or four things you wrote yourself — a memo, a long email, a post you liked. Unpolished is better than polished."

If the user has nothing written, capture voice by interview instead: ask four questions they care about and transcribe how they answer. Speech is closer to real voice than most business writing.

What to extract

Write to ${SALIENCE_HOME:-~/.claude/salience}/voice.yaml:

• Sentence rhythm — average length, variance, whether they use fragments
• Opening habits — how they start: claim, question, story, concession
• Signature constructions — patterns that recur across samples
• Vocabulary — words they reach for, and the register they hold
• Argument shape — do they lead with the conclusion or build to it
• Concession behavior — how they handle the counterargument
• Humor — present or not, and what kind. Dry, self-deprecating, none
• Hedging tolerance — how much qualification is natural to them
• Formatting — lists vs. prose, paragraph length, emphasis habits
• Anti-patterns — things they visibly never do

Anti-patterns matter as much as patterns. A person who never uses exclamation points and never opens with a question has told you two firm rules.

Confidence zones

A flat voice profile makes a person sound equally certain about everything, which is the fastest way to sound fake. Map expertise into three registers and record which topics sit where:

For an executive this is the difference between credible and grandiose. Writing about a core discipline in exploration voice reads as falsely modest; writing about something genuinely new in full-authority voice is the single most damaging voice error available, because the audience most likely to notice is the audience being targeted.

Record the zone per topic. salience-content reads it when choosing how to pitch a claim.

Validate before saving

Write the same short passage twice — once in the captured voice, once deliberately off it — and show both:

This sounds like you:
  "The measurement layer broke before the marketing did. Everyone argued about creative for
   two quarters."

This doesn't:
  "In today's evolving landscape, it's crucial to leverage data-driven insights to unlock
   marketing potential."

Then ask: "Does the first one sound like you when you're not overthinking it? What's off?"

The contrast is what makes the test work. Shown alone, almost any competent passage reads as plausible; shown against a wrong version, people identify the mismatch immediately.

Source every anti-pattern to evidence. "You never used 'leverage' across eleven samples" is a finding. "Avoid corporate jargon" is a guess wearing a finding's clothes.

An unvalidated voice profile is a guess that will silently distort every future output. If the user says it is close but off, ask what specifically is off — that answer is usually the most valuable line in the whole profile.

Check your own work first

Before showing the profile, answer these honestly. The same evidence discipline that governs career facts governs voice extraction:

• Are the signature phrases actually in the samples, or did I infer them from tone?
• Does the anti-pattern list name specific words, or vague categories?
• Do the two validation passages differ in a way a reader would actually notice?
• Are the confidence zones mapped to named topics, or generic?
• Could someone else write in this voice from this document without asking a follow-up question?

Flag the gaps rather than papering over them: "The anti-pattern list only has two entries, which isn't enough to constrain anything. Send me two more samples or tell me three phrases you'd never use."

Enforce

Against the voice profile

Check rhythm, openings, vocabulary, argument shape, and formatting. Flag deviations with the specific fix.

Against AI tells

Strip regardless of what the voice profile says. See ../salience-profile/references/language-scan.md for the full list. The core:

Vocabulary: leverage (verb) · delve · unlock · harness · foster · streamline · robust · comprehensive · fundamentally · seamless · elevate · empower · navigate (metaphorical) · landscape (metaphorical) · testament to · realm · tapestry

Constructions: "It's not just X, it's Y" · "In today's fast-paced…" · "I'm excited to announce…" · rule-of-three everywhere · uniform paragraph lengths · closing by restating the opening · rhetorical question openers used as a formula

Punctuation: em-dash rhythm the person does not otherwise use · decorative arrow bullets · emoji in executive copy · exclamation points

Never rewrite an engineered hook

A hook that another module deliberately constructed — for tension, specificity, or a fold constraint — is not subject to naturalness editing. Rewriting it to sound more like the user destroys the structure it was built for, and it is the most common way a voice pass makes a draft worse.

Enforce voice on the body. Leave a hook alone unless it contains a forensic tell or an unsupported claim, and if it does, say so and hand it back rather than quietly smoothing it.

Tiered, not absolute

Not every flagged pattern is wrong. Some people genuinely write with em dashes and rule-of-three lists, and stripping those makes the text less like them, not more.

When the voice profile contradicts a general rule, the voice profile wins. The goal is sounding like the user, not sounding like a style guide.

Reporting

Voice check — 3 changes

  "Leveraging AI to fundamentally transform demand generation"
→ "Using AI where it actually removes work"
  Two forensic tells. You don't write "leverage" in any of your samples.

  "It's not just about tools — it's about outcomes."
→ "The tools were never the problem."
  Formulaic construction; your samples make this move by flat assertion.

  Four paragraphs, all 3 sentences.
→ Broke the third into a one-line paragraph.
  Your writing varies hard between long and very short. The evenness read as generated.

Explain the meaningful ones. Silent correction teaches nothing and the same phrasing returns next draft.

Boundaries

• Voice is not a licence to overstate. Confident phrasing must still rest on a real fact — the evidence contract outranks the voice profile.
• Never fabricate a personal anecdote to sound authentic.
• Multiple registers are normal — a board memo and a LinkedIn post differ. Register flexes; identity does not.

References

• references/voice-profile.md — the schema, the capture interview, worked extraction
• references/ai-tells.md — full pattern list with tiers and replacements