Keeping Your Voice When You Write With AI

October 5, 2026 · 5 min read

The short answer: keep AI in the role of draft producer and stay yourself in the role of editor. Generated text enters your piece only after you have decided it belongs, and the deciding — what to keep, what to cut, what to refuse — is where voice actually lives. The workflow below works in an ordinary chat interface like SentX's, and it scales from a marketing page down to a three-line poem. It carries one material condition: under SentX's published privacy policy, submissions are not confidential, and retained interactions can enter shared memory. The workflow therefore assumes you write only with material you could stand behind in the open.

What does "your voice" mean in practice?

Not a mood you hope the model guesses. Operationally, voice is a set of choices you can name: the words you reach for and the ones you avoid, the sentence lengths you tolerate, the images you find true and the ones you find cheap, the subjects you would never hand to a machine. Write those down. A brief of four or five lines — subject, audience, mood, framing, deliverable, plus one or two sentences showing how you would open it yourself — gives the model something concrete to work from, and more importantly gives you a standard to judge its output against. An example, offered as illustration rather than instruction: "A 400-word piece on why small businesses ignore their own support tickets. Reader: operations managers. Tone: direct, no corporate language. I open with the ticket itself, quoted, not with a statistic." That brief is checkable. "Write something engaging" is not.

What is the order of operations?

1. Write the brief first, in plain language, before any generation happens.
2. Settle strategy in conversation: sketch two or three possible angles, pick one, and only then ask for the draft.
3. Work in single-variable passes — change the tone, or the length, or one central image, not all three at once, so you can see which change did what.
4. Review each pass against the brief, not against how it feels in the moment.
5. When a version earns approval, save three things outside the chat: the text you chose, the prompt that produced it, and the brief.

Steps two and five are the ones people skip. Abandoning a direction in a message costs nothing; abandoning it in a finished draft costs the draft. And the saved prompt-plus-brief pair is what makes the collaboration reproducible next time, while the saved text is the only artifact that actually matters.

How much of the draft should you keep?

The two failure modes are total adoption — pasting the whole draft and calling it done — and total rejection, which wastes the tool. Selective revision sits between them: take the sentence that lands, rewrite the one that almost lands, delete the rest. There is no honest ratio. Some pieces will keep most of a good draft; others will keep one image out of ten pages. Authorship was the choosing either way.

Give the review step observable criteria so it stops being a feeling. Does the draft say what the brief says, and nothing the brief does not authorize? Are there phrases you would never write — each one is data about your voice, worth feeding back into the next pass? Does any compressed line — a title, a summary, an intro — promise more than the detail behind it delivers? Compression invites overclaiming, and the model will not reliably flag its own exaggeration. That check is yours.

How should haiku be handled differently?

Haiku is the case where people most often hand the whole piece to the machine and walk away satisfied with whatever comes back, and the form hides that badly. What the form actually asks, per the Academy of American Poets' glossary, is modest: a brief verse, rooted in Japanese tradition, typically built around a fleeting moment — often natural, often seasonal. The familiar three-line, five-seven-five syllable pattern is presented there as the conventional English rendering, and it is one: the glossary does not claim all English haiku must follow it. So treat 5/7/5 as a constraint you may impose on the generator, not a law the form imposes on you. Ask for haiku with the same brief discipline — a specific moment, not "nature"; "in the 5-7-5 pattern" if you want the shape, omitted if you don't — and hold the output to what the syllable count cannot enforce: a concrete, particular image, and a suggestion that outruns the statement. A generated line that scans perfectly and says the wind is gentle has failed the form while passing the meter. The meter is the easy part; the seeing is yours to supply.

If you already write haiku, invert the exercise: write three first, then ask the model to respond to them — extend a season word, offer an alternative image for the weakest line. Now it works inside your voice instead of replacing it.

What are you giving away?

Stated plainly, because SentX publishes it and SentX is the publisher of this guide: its privacy policy says submissions are not confidential; representations of retained interactions can enter shared memory and influence other users' interactions; and deleting an account does not promise removal of material already incorporated into shared memory or model weights. It simply sets the boundary. Unpublished work you intend to protect, client names, personal correspondence — keep them out of the prompt and the brief alike. Everything else, most of what writers actually draft, fits inside the boundary.

None of this requires refusing the tool, and the cost side is small — a few lines of brief, a conversation before the draft, a checklist at the end. What it protects is the part of writing that does not scale: the seeing, the cutting, the refusal.