How to Write Naturally with AI — co-write, then claim the voice.

Step 1: Set the angle, then let the model propose structure

Start the collaboration by giving the model something only you know: the angle, the audience, and the one claim you want a reader to leave with. Then ask it to propose three competing outlines, not one. If your piece is "what changed after we moved our standups to async video," you might get a chronological outline, a problem-and-fix outline, and a contrarian one that argues async failed first. Pick the spine you believe and discard the rest. You are using the model as a structure partner that surfaces shapes you would not have reached for, while the decision about which shape is true stays with you. That single choice is the first place your voice enters the co-written piece.

Step 2: Co-draft section by section, never the whole piece at once

Drafting together means working one section at a time, with you steering each turn. Hand the model your chosen outline plus two or three of your own raw sentences for the section, and ask it to continue in that register, not to start fresh. Read what comes back as a proposal, keep the line of reasoning, and mark every phrase that sounds like a model rather than you. The reason to co-draft in small sections is control: a full-document generation hands the model the connective tissue between paragraphs, and that connective tissue is exactly what a reader hears as machine cadence. Section by section, you stay the lead author and the model stays the second pair of hands.

Step 3: Take the ownership pass and re-say every borrowed line

This is the step that makes a co-written draft read as natural, and the one most people skip. Go through the merged draft and re-say each sentence the model contributed in your own words, out loud first if it helps. Not edit, re-say: if you cannot explain why the sentence is there or restate it without looking, cut it and write the point yourself. A 800-word co-draft usually needs 20 to 30 minutes of this ownership pass, and it is where the collaboration either succeeds or quietly fails. The goal is that by the end, no line survives that you could not defend as a choice you made rather than output you accepted.

Step 4: Use the model for narrow fixes, with you holding the pen

Once the ownership pass is done, the collaboration narrows. Ask for three word alternatives where one is not landing, a grammar check on the finished piece, or whether a sentence reads clearer with the modifier at the front or back. Each of these keeps you holding the pen while the model checks your work. What it should not do at this stage is propose a smoother version of a paragraph you have already claimed, because a smoother version reintroduces the cadence you just removed. Accept changes that fix a real mistake and reject changes that quietly hand a sentence back to the model.

Step 5: Scan with TextSight to confirm the handoff worked

Paste the co-written draft into TextSight at app.textsight.ai. The scan returns an overall score, a sentence-level highlight map, and a Plagiarism Risk score in the same pass. In a healthy co-writing loop the highlights cluster on exactly the lines you let the model keep without re-saying, which is the cleanest possible feedback: it tells you which turns of the collaboration you under-owned. Re-say those specific sentences and rescan. A co-drafted piece that has been through a real ownership pass usually lands in the natural band, with scattered residuals rather than the dense clusters that mark an unedited generation.

1. Replace the model's generics with your specifics

Models default to safe, generic claims because that is what averaged training data produces. Where the co-draft says "many writers struggle with editing," swap in something only you would know: "the three freelancers in my Tuesday writing group all skip the second pass." Where it says "the meeting was unproductive," write "we spent 45 minutes debating font choices for a slide deck that would not exist in six months." The model cannot invent these because it never lived them; it only synthesises from training data. Each specific you trade in is a sentence the model can no longer claim authorship of. Five concrete swaps usually flip a section from machine-flat to recognisably yours, even before the rhythm work.

2. Break the even rhythm the model defaults to

Co-drafted prose tends toward uniform sentence length, because the model optimises for clarity and lands almost every sentence between 16 and 22 words. Detectors call that low burstiness, and it is the most common reason a human-led draft still flags. As you take the ownership pass, deliberately plant one short sentence (under 8 words) and one long sentence (over 28 words) in every paragraph the model touched. The contrast carries voice. Read each reworked paragraph aloud; if every sentence takes the same breath, the model's cadence is still running underneath yours. This is the cheapest fix in the whole loop and the one that moves the score most reliably.

3. Inject lived detail the model could never supply

The one thing no co-writing partner can contribute is your own experience, so it is the fastest way to take a section back. The bar is lower than you think: a one-sentence reference works. "I tried this workflow on a newsletter draft last Sunday" or "the email that taught me this came from a reader who works in pharmaceutical regulation" is enough. Drop one personal reference into every section the model helped draft. It is the line the model could never have produced on your behalf, because it has never lived a Tuesday, and a reader feels the shift from synthesised to remembered immediately.

4. Re-say each line the way you would say it aloud

The final move is the register correction. Take each sentence the model kept and re-say it as if you were explaining the point to a friend at a coffee shop: use contractions, trust short sentences, and strip the scaffolding the model leans on ("in today's rapidly evolving landscape," "it is important to note that," "this article will explore"). The test for a co-written line is simple. Would you say it out loud to a friend in those words? If the answer is no, it is still the model's sentence wearing your byline, and it needs one more pass before it is genuinely yours.

Light: for a line you nearly owned

Light keeps the wording close to what is already on the page, which makes it the right mode for a sentence you mostly re-said but the scan still flags. Run it on the two or three leftover lines from your ownership pass, never on a section. Light typically moves a flagged sentence score by 15 to 25 points without changing the meaning. If the result drifts even slightly from the point you were making, reject it and re-say the line by hand instead. At this stage the rewriter is finishing a handoff you almost completed, not taking a turn back from you.

Standard: for a paragraph stuck in the model's rhythm

Standard rewrites more aggressively and fits a co-drafted paragraph that still flags after you re-said it, usually because the model's even rhythm survived your pass. It breaks that rhythm more reliably than Light, at the cost of a slightly less faithful echo of your phrasing. Run it on one paragraph at a time, never the whole piece, and treat the output as a fresh proposal in the collaboration: read it aloud and keep only the lines that sound like you would actually say them.

Maximum: rarely right after a real ownership pass

Maximum is built for heavy AI drafts where the prose was never claimed in the first place, which is the opposite of a properly co-written piece. Running it on a draft you co-wrote and then owned usually makes it worse, because it discards the voice you just worked into the lines. The two times Maximum earns its place on co-written work are translation polish (the ideas and structure are yours but the English needs heavy work) and emergency salvage on a deadline you will not make otherwise. In both cases the ownership pass moves to afterwards, and a manual read-through is non-negotiable.