Key Takeaways
- Describing your tone ("casual but professional") gives a model nothing to act on; pasting real writing samples and asking it to describe measurable patterns does.
- A banned-word list works better than a tone description — and the list that matters most is the one only you can write: words and phrases you personally never use.
- Comparing a piece that measurably worked against one that didn't gives a mechanical diagnosis a model can act on; "make it better" doesn't.
- None of this is one-time setup — a new conversation doesn't remember your samples or banned list, so the fix is saving the reference once and re-pasting it, or moving to a system that holds it as standing context.
Every model defaults to the same voice: hedged claims, symmetrical three-item lists, transitions that restate the sentence before them. That voice is a statistical average of everyone who's ever published online, and averages don't sound like anyone in particular — least of all you. Fixing that isn't a prompt you paste once and forget. It's a short, repeatable process you run before you ask for a draft, and most people who complain that AI writing sounds off have never run any part of it.
This is the playbook version. For the mechanism behind why the default voice sounds the way it does, see Why AI Output Is Generic, and What Actually Fixes It; for the model-specific breakdown, Why Does ChatGPT Sound Generic? covers that angle. Here's what to actually do, in order.
Why doesn't telling AI to "write in my tone" work?
"Casual but professional" and "friendly but direct" are labels you already agree with about yourself — they carry no information a model can act on, because everyone thinks they sound casual but professional. What it can act on: sentence length, punctuation habits, contraction rate, where you place the verb, what you repeat without noticing. Those are measurable patterns, and measurable patterns are instructions.
Paste in three to five things you actually wrote start to finish — an email, a Slack message, a proposal paragraph — and ask the model to describe the pattern, not judge it. A prompt that works:
"Here are four emails I sent this quarter. Tell me my average sentence length in words, whether I use contractions, how often I open a sentence with 'And' or 'So,' any word or phrase I use more than once, and whether my sentences run short-declarative or longer with clauses. Describe the pattern — don't evaluate the writing."
Renata runs finance at a 30-person freight brokerage in Charlotte and ran this against four vendor-negotiation emails from the past quarter. The pattern that came back: 14-word average sentence, zero contractions, the ask stated in sentence one, no instance of "I think" or "I believe" anywhere. She pastes that description into every finance-memo prompt now, and the drafts stopped reading like a template pulled off a course.
Why does a banned-word list work faster than describing your tone?
Models default to a specific inventory of words and habits: "leverage," "seamless," "robust," "delve," "moreover," rhetorical questions, hedges like "could potentially," and lists that land suspiciously often on exactly three items of matching length. Naming these explicitly removes them faster than any amount of tone coaching.
But the banned list that matters most is the one only you can write: words and phrases you personally never use. That's different for every writer, and it's the half of the list people skip.
Marcus runs an 8-person branding studio in Austin and writes every client-facing proposal himself. His banned list has 22 entries. He added "elevate" after the fourth proposal draft in a row used it to describe a logo refresh — a word he's never once used to describe his own work in six years of writing these documents. The list lives in a text file he pastes at the top of every drafting prompt: never these words, never open with "In conclusion," never exactly three items in a list.
How a Hit-and-Miss Comparison Fixes Generic AI Writing
Find one piece of your writing that measurably worked — a cold email with a real reply rate, a LinkedIn post with actual comments, not likes — and one on a similar topic that didn't. Paste both in and ask for the mechanical difference, not a verdict.
Lauren owns a 22-person landscaping company outside Columbus and writes her own renewal-notice emails. She compared one that got a 60% same-day response against one that got almost none. The model's answer wasn't "the first one is better written" — it named the actual difference: the working version asked a single question in the first line, and the flat version buried the ask in paragraph three behind two sentences of throat-clearing context. That's a fixable, specific finding. "Make it better" never produces one.
The prompt that gets a diagnosis instead of a compliment: "Post A got 40 comments from people in my industry. Post B got 3 likes and no comments. Name three specific mechanical differences — structure, sentence length, word choice — not which one is better."
Why doesn't "make it sound more like me" work as feedback?
"Make it sound more like me" is exactly as useless as "casual but professional" — it's a label, not a location. The model has nothing to correct against, so it rewrites toward the same statistical average it started from. Feedback that actually changes output points at one place and one problem:
"Cut the second sentence in paragraph 3 — it restates paragraph 2." "This list has three items of matching length and structure, break the rhythm on the third." "Sentence 4 hedges with 'might potentially' — I don't hedge; make it a direct claim or cut it."
One note, one revision, then look at what came back before sending the next note. Stack five vague instructions into a single message and the model averages your notes the same way it averaged its training data — which is the exact problem you're trying to undo.
Do you have to redo your AI voice calibration every new session?
Here's the part that doesn't get said enough: none of this is one-time setup. A new conversation doesn't remember your sample sentences, your banned list, or Tuesday's hit-and-miss diagnosis. Open a fresh chat and you're recalibrating from zero unless you're the one carrying the reference over. That recurring cost — not any single technique above — is the actual complaint sitting underneath "why doesn't AI writing sound like me."
The workaround that makes the cost bearable: save the sentence-pattern inventory, the banned list, and your hit/miss notes in one document, and paste the relevant pieces into every new writing session as fixed context. It doesn't remove the setup. It removes rebuilding the setup from memory every single time.
When does manually recalibrating AI writing stop being worth the time?
Reopening a notes file, re-pasting four sample emails, re-explaining a 22-word banned list — every session, for every piece of writing — is exactly the kind of tax that gets old once you're doing this daily instead of occasionally. That's the part Kuvai's approach automates: a teammate keeps your voice and style reference inside its Company Context once, instead of needing it re-supplied at the top of every conversation.
It doesn't skip the work described above. You still hand it the real samples, the real banned list, the real hit-and-miss comparison — honestly, the same way Renata, Marcus, and Lauren did. The difference is what happens after: it accumulates your context instead of starting blank each time, so the fifth email of the week doesn't cost you the same fifteen minutes of setup as the first.
Want AI writing that's actually grounded in your voice, not a generic default? Sign Up for Free — no credit card required, free to start, cancel anytime.