Everyone can spot AI writing now. The em dashes, the "delve into," the relentless competence, the way every paragraph could have been written by anyone about anything. Readers call it sludge even when they can't name why. Editors reject it. And writers who use AI worry, correctly, that they're sanding down the very thing that makes their work theirs.
The good news: the problem isn't that you use AI. It's how you use it. AI is excellent at tasks that support your voice and terrible at tasks that replace it — but the default way people prompt reverses that, asking AI to generate and then trying to humanize after.
This guide flips the workflow: extract your voice first, set guardrails, then use AI only where it strengthens your style instead of replacing it.
Why AI flattens voice
AI models are trained to produce the most probable next word. Probability favors the middle — the sentence that offends no one, surprises no one, and sounds like everyone. Your voice, by contrast, lives in the improbable: the specific word you'd choose, the rhythm you'd break, the thing you'd notice that a model would skip as low-probability.
Three mechanisms flatten it:
- Regression to the mean. Ask for "a professional introduction" and you'll get the average of every professional introduction the model has seen. Your idiosyncrasies are statistical outliers — the model smooths them away.
- Helpfulness over honesty. Models are tuned to sound confident and polished. Your voice might be tentative, jagged, funny, or spare. Polishing it out isn't an improvement — it's a different writer.
- No stakes. You choose words because they matter to you. The model chooses words because they are likely. Without intent, there's no voice.
Understanding this tells you where AI helps and where it hurts. Generation from scratch hurts. Structured editing, constraint-based rewriting, and analysis of your existing text help.
Step 1: Extract your voice before you automate anything
You can't protect a voice you haven't articulated. Don't ask AI to "write in my style" before you've shown it what that style is.
The extraction pass. Collect 2,000 to 3,000 words of your writing you like — not your best sentences, but the ones that feel most you. Then run this:
Here are three samples of my writing: [paste samples]. Analyze my voice across these dimensions: sentence length and variation, vocabulary level, use of metaphor vs literal description, humor or seriousness, how I open and close paragraphs, and any recurring tics. Give me a 10-bullet style guide in my own terms, with a short example phrase for each bullet illustrating the trait.
Save that guide. It becomes the constraint for everything after. Our Write in My Style prompt does a compact version of this — paste your samples and it mirrors your patterns without copying content.
The contrast check. To sharpen the guide, ask the inverse:
Rewrite this paragraph I wrote: [paragraph] — first, in the most generic AI style possible, then in my style as described above: [past style guide]. Put the two versions side by side and tell me the 4 biggest differences.
What changes is your voice. What's left when you subtract sludge is style.
Step 2: Set guardrails for every AI task
With a guide in hand, you can use AI without anxiety. Attach guardrails to every prompt.
The voice guardrail template:
Use my style guide: [paste 10-bullet guide]. Follow it strictly. If you must deviate for clarity, flag the deviation in brackets and explain why. Never add em dashes, "delve," "unlock," "elevate," or semicolons unless they appear in my samples.
That last line matters more than it seems — it bans the model's favorite tells. Customize the ban list to your own pet peeves.
Our Match a Brand Voice prompt formalizes this into a reusable instruction set. Build it once, then append it to any writing task.
When to use vs. not use AI:
| Good for AI | Bad for AI |
|---|---|
| Analyzing your drafts for repetition and filler | Writing first drafts that need your experience |
| Offering tighter phrasings you choose between | Choosing the final phrasing for you |
| Finding where voice slips (inconsistency detection) | Inventing voice from nothing |
| Research, structure, and summarization | Sentences that carry emotion or vulnerability |
| Copy-editing with annotated reasons | Unannotated rewrites that change meaning |
The rule: AI proposes, you dispose. If the model makes a word choice you wouldn't, it's wrong even if it's grammatical.
Step 3: Editing without flattening
Editing is where most voice is lost. Someone — or something — "cleans up" your draft and returns it smoother, more correct, and less you. Fix this by demanding annotated edits.
The annotated edit. Our Line Edit With Editor's Notes prompt was built for exactly this:
Edit the following text for clarity — cut wordiness, fix grammar, break long sentences — but do not change my meaning, voice, or key terminology. For each significant edit, add a brief [Editor's note: reason] so I understand the change.
The annotation is the point. When you see why something changed, you can accept, reject, or adjust. Without it, you rubber-stamp sludge.
The humanize-last trap. Many writers generate with AI and then ask it to "make this sound more human" or use a Rewrite as a Human pass. This adds surface variation — contractions, idioms, sentence fragments — but the thinking is still generic. Readers sense it.
Instead, humanize first. Write the messy, specific, slightly embarrassing first draft yourself — the one with the wrong words and the real story. Then let AI do constrained cleanup:
Here's my rough draft: [draft]. My style guide: [guide]. Clean up grammar and tighten sentences over 28 words. Do not add new ideas, metaphors, or transitions. If a sentence is unclear, ask me rather than rewriting it.
That last instruction — ask rather than rewrite — prevents the model from inventing when it should be uncertain. It's the most important sentence in the prompt.
The clarity vs. voice tension. Sometimes AI suggests a clearer version that is genuinely better writing but less you. Our Rewrite for Clarity prompt navigates this by offering options on a spectrum:
Use it to get three versions of a paragraph — from closest to your original to most revised for clarity — so you can see the tradeoff and pick your spot, not the model's.
Step 4: Study voice without stealing it
Learning from authors you admire is different from having AI write as them. Do the second carefully.
Style analysis, not style forgery. The Write in the Style Of prompt is designed for this distinction — it imitates a named author, then breaks down the five techniques behind the style so you can borrow craft, not sentences:
Write a 300-word passage about [topic] in the style of [author]. Then break down what you did: identify 5 specific techniques that define this style — sentence structure, word choice, pacing, metaphor, narrative distance. Finally, show me how to apply 2 of these techniques to my own voice without copying.
The breakdown matters more than the imitation. You want the principle ("Joan Didion uses short declarative sentences after a long wandering one to create drop"), not the paragraph.
Your voice evolves — retrain. Every 10,000 words or so, re-run the extraction pass with fresh samples. Voice isn't static. If you don't update the guide, you'll be protecting a version of yourself you've outgrown.
A workflow that keeps you the author
Put it together as a repeatable process:
- Draft alone. Messy first draft, your words, no AI in the room. This is where voice originates.
- Extract / update guide. Run the extraction prompt if you haven't in a while.
- Annotate-edits only. Use Line Edit With Editor's Notes with your guide attached. Review every note and reject anything that isn't you.
- Pressure-test with imitation. Use Write in the Style Of to study how other voices handle your material, then steal one technique at a time.
- Final human read-aloud. Read the draft out loud. Any sentence you stumble on or wouldn't say out loud gets rewritten by you, not the model.
AI is a better editor than writer. A better analyst than author. Use it where it's strong and keep the sentences that make you you for yourself. That's not romantic — it's practical. Generic writing is now free and infinite. Only specific, voiced writing stands out.
Explore the full set: Write in My Style, Rewrite as a Human, Rewrite for Clarity, Line Edit With Editor's Notes, and Match a Brand Voice in the Promptigo library.