We've already written about why everything ChatGPT writes sounds the same: a blank prompt box has no memory of your positioning or your opinions, so it averages everyone else's. If you want the mechanism, start there. This post is the other half — the how. Because the fix isn't better prompts or a more expensive model. It's a voice layer: a deliberate set of inputs that sits between the model and the page, so that every draft starts from who you are instead of from the median of the internet.
What does "sounds like you" actually mean?
It means a regular reader could not tell which paragraphs you typed and which were drafted for you — indistinguishable from you, not undetectable as AI. That distinction matters, because the two goals pull in different directions. Chasing "undetectable" means optimizing against detection tools, which is a losing game aimed at the wrong judge — and an unnecessary one, since Google doesn't penalize AI content for being AI-made, it penalizes content that's unhelpful and interchangeable.
The judge who matters is a reader who already knows you — from your LinkedIn posts, a sales call, a podcast. If that person reads the piece and hears you, you've won. If they hear a competent stranger, no detector score will save you. So the target isn't hiding the tool. It's making the output carry so much of your position and your material that it couldn't have been written about anyone else, by anyone else.
The three inputs a voice layer needs
Strip any founder's distinctive writing down to parts and you find three ingredients. A voice layer is just those three, captured explicitly enough that a model can use them:
Position
Who you're for, who you're against, and the takes you'd defend under pressure. "We think weekly consistency beats viral swings" is a position. "Quality content matters" is wallpaper.
Voice mechanics
Your sentence habits, the vocabulary you actually use, a never-say list ("game-changer," "delve," whatever makes you wince), and the analogies you reach for on calls.
Lived material
Customer stories, real numbers, decisions you made and why. This is the input no model can hallucinate correctly — which is exactly what makes it unmistakably yours.
Most "make AI sound like me" advice stops at the second ingredient — tone, style, formatting. That's why it half works. Tone without position produces a well-dressed nobody; tone without lived material produces confident essays with no fingerprints on them. You need all three, and the third is the one generic drafts can't fake.
How do you capture your voice for AI?
Two ways: maintain a living voice document and example bank you paste into every session, or capture the context once in a system that persists. Both work. They differ in how much of the maintenance lands on you.
The manual version. Build a voice document — one to two pages, not a brand bible. Your one-sentence audience, five to ten positions phrased as sentences you'd actually say, your never-say list, three analogies you use constantly. Alongside it, an example bank: four to six pieces of writing that genuinely sound like you — your best LinkedIn posts, a sharp email, a transcript excerpt where you explained your product well. Paste both at the top of every AI session before you ask for a draft, and add the specific lived material for that piece: the customer story, the number, the decision.
Be honest about the failure mode, because it's predictable: decay. The document is accurate the week you write it and drifts every week after. Your positions sharpen on sales calls, your never-say list grows, your best recent writing never makes it into the example bank — and every session starts from scratch, so anything you fixed in last week's chat is forgotten in this week's. Manual voice layers don't fail loudly; they just quietly fall out of date until the drafts sound like the founder you were six months ago.
The systematized version. The alternative is to capture the context once and have it applied to every draft automatically. This is the part of the problem FirstOrg was built around: an onboarding interview extracts positioning, opinions, and voice instead of handing you a blank form, and Deep Lattice keeps that context — plus what it learns from your edits — persistent across everything the engine produces, so the voice layer compounds instead of decaying. That's the honest pitch: not writing that's magically you, but your context, captured once, present in every draft, and updated as you evolve — the maintenance job the manual version leaves on your calendar, handled.
The human pass that stays
No voice layer removes the final human pass, and you shouldn't want it to. This is the pass that turns "authored by your context" into "authored by you" — and it's minutes per piece, not a rewrite. Three moves:
- Kill the hedges. Models soften by default. Delete "arguably," "in many cases," "it's worth noting" — every qualifier you wouldn't say out loud.
- Sharpen the claims. Where the draft says "this can be an effective approach," write what you actually believe: "this works" or "this fails, here's when."
- Add the detail only you know. One customer name (with permission), one real number, one "we tried this and reversed it" — a single concrete detail does more for authenticity than any style instruction.
If you're uneasy about publishing words you didn't type, we've addressed the ethics head-on in whether using a ghostwriter is cheating — the short version is that authorship lives in the point of view, and this pass is where you stamp it.
How do you test whether it sounds like you?
Run three checks on every draft: read it aloud, ask whether you would defend each claim on a sales call, and try the swap test. Each catches a different failure.
The read-aloud test catches voice-mechanics failures. Read the piece out loud, literally. Any sentence you stumble on, or would never say to a human, gets rewritten in the words you'd actually use. Your ear knows your voice better than any rubric.
The sales-call defense test catches position failures. For every claim in the piece, ask: if a prospect quoted this back to me on a call, would I stand behind it — with reasons? If the honest answer is "well, it's more nuanced," the piece is carrying a position you don't hold. Fix the claim or cut it.
The swap test catches lived-material failures, and it's the brutal one: could your closest competitor put their logo on this piece and publish it unchanged? If yes, nothing in it is yours yet — no story, no number, no take a rival couldn't co-sign. A piece that passes the swap test is, by definition, content that doesn't sound like AI, because it couldn't have come from anywhere but you.
That's the whole system: three inputs captured, one place they persist, a minutes-long human pass, three tests before publish. The model drafts. The voice layer makes it yours.