Why a draft doesn't read like ChatGPT

Three separate mechanisms do this work. A measurement of your own writing catches what actually repeats in it. A checker runs after the model and can refuse to accept what it wrote. A hard boundary limits what the extension is even allowed to read in the first place. None of the three is a prompt asking the model to "sound human" - each is code you can read, named below next to the file it lives in.

What the system reads about you

Two tables hold what Pitchbox knows about you specifically: operator_profiles (your persona - handle, headline, about, experience) and operator_voice_samples (your own recent posts). Both are populated by the Chrome extension's content script reading whatever your own browser already rendered on your own /in/<slug> and recent-activity pages (extension/src/content/linkedin-profile-capture.ts), and both are editable by hand from the Companion section of the dashboard - the persona at /companion, the voice samples (excludable one by one) and the derived summary at /companion/voice.

The boundary that makes this safe: no code path in this product ever issues a request toward linkedin.com. The extension reads the DOM already in front of you and posts what it found to Pitchbox's own backend - the profile-capture script's own comment spells out why the request path is even named apart from the word "linkedin" (OPERATOR_PROFILE_PATH, same file), so a static scanner can tell "posts to our backend" from "fetches LinkedIn" on sight. That scanner is real, not a promise: pnpm run test:linkedin-compliance (tests/compliance/linkedin-boundary.test.ts) parses the built extension and fails on six things - a fetch toward linkedin.com or licdn, a document.cookie, chrome.cookies, localStorage, or sessionStorage read inside a LinkedIn content script, a synthetic click or submit on a LinkedIn element, an alarm reachable from LinkedIn code, a network call anywhere in the LinkedIn platform directory, and the LinkedIn host permission turning into a blanket grant instead of the optional one it has to stay. It is a required CI check, and it inspects the real manifest and source, not a checklist someone filled in.

How the voice profile is derived

shared/src/assist/voice-profile.ts is a measurement, not a model call. It reads your voice samples, your sent messages, your sent drafts and your project templates, and counts what actually repeats:

• Sentence length, and how much it varies from one piece of writing to the next.
• Paragraph shape.
• Punctuation habits.
• How often you write in the first or second person.
• How often you hedge.
• Which words you reuse.
• The split between English and Italian.

Every one of those is a countable property of text that already exists, computed with no I/O and no model in the loop.

The floor is explicit: MIN_ITEMS_TO_DERIVE is 3. Below that, the function returns nothing measured, because two posts sharing an opening word is a coincidence, not a habit - a smaller corpus can't tell the difference and the code doesn't pretend otherwise. Below the floor, shared/src/assist/voice-defaults.ts supplies a fallback instead: a short, named set of rules ("open with a short first sentence," "no tricolons," "plain words") that is stated as a default in its own text, never phrased as if it came from reading anything you wrote. shared/src/operator-voice-profile.ts's resolveVoiceProfile is what keeps the two from blurring together - every axis it hands back is labelled measured or default, never both, so nothing downstream can present a guess as a fact about you.

The house style, and why it is code

shared/src/house-style.ts holds the prose every playbook carries verbatim, under the heading ## House style: write like a human. shared/src/style-check.ts is the part that actually runs against what a draft came back with. It catches, among other things:

• Em dashes, curly quotes and curly apostrophes, the single-character ellipsis, non-breaking spaces - e.g. word — word becomes word, word.
• Filler openers - "Great post!", "Thanks for sharing".
• Puffery words - leverage, unlock, delve, game-changer.
• Wrap-up closers - "at the end of the day", "hope this helps".
• The "not just X but Y" construction.
• A rhetorical question as the opening sentence.
• Tricolons (three items closed with an Oxford comma and "and").
• Hashtag stacks and emoji used as bullet points.

There are two kinds of finding, and they are handled differently. A character-level finding (an em dash, a curly quote) is repaired mechanically by applyMechanicalRepairs - a string substitution, no model call. A structural finding (a tricolon, a rhetorical opener) can't be fixed that way, so enforceHouseStyle sends the draft back to the model exactly once, naming the exact span that failed, and checks the result again. A finding still present after that round trip is never silently dropped: it travels with the draft so a human sees it rather than a cleaned-up draft that quietly still has the tell.

What you can steer

• Tone. Settings → LinkedIn assist sets a default register for every suggestion, chosen from six presets:
  • Match the room - the default, mixing the post's own register with your voice.
  • Professional.
  • Plain.
  • Warm.
  • Technical.
  • In my own words - a free-text note you write yourself.
• Retune, per suggestion. The panel itself offers Drier, Warmer or Shorter on the draft currently on screen, without touching the stored tone setting - it's a one-off request on that call, not a preference.
• Voice samples. Any post under /companion/voice can be excluded from the corpus without deleting it, so a capture you don't want feeding the measurement doesn't just come back on the next page load.
• The derived summary itself. You can hand-edit it, and a manual edit survives the next refresh until you explicitly reset it back to derived (resetVoiceProfileToDerived in shared/src/operator-voice-profile.ts).

What it will not do

It does not send anything. The extension writes drafted text into the composer already open on the page (insertComposerText in extension/src/content/linkedin-comment.ts) and copies it to the clipboard as a fallback - you press LinkedIn's own submit button, always.

It does not imitate anyone but you. Every corpus item the voice measurement reads is scoped to your own organization - your own samples, your own sent messages, your own sent drafts.

It does not claim to know your voice from nothing. Below the three-item floor it says so in its own text rather than guessing: "No writing on file yet, so this is a default style, not a measurement."