AI newsletter writer that keeps your voice

A newsletter is the one channel where people opted in specifically to hear from you. That makes it the worst possible place to sound like a content mill.

Why it matters here

Newsletter readers notice drift faster than any other audience, because they read you repeatedly. Voice consistency across issues is the whole asset, and it is exactly what degrades when you draft with a general-purpose model and a tone description.

How ScriptGrain does it

  1. Feed in back issues you were happy with. The voice profile is built from what you have already sent.
  2. 45 measured attributes capture the things a style guide never manages to write down: your discourse markers, your paragraph openers, how declarative you are versus how much you hedge.
  3. Each issue is drafted against that profile, so issue forty still sounds like issue one.

Newsletters default to around 500 words with an outline you approve first, so structure is settled before a word of draft is written.

What an issue in your voice looks like

Start with what a reader actually sees. Not "tone". Rhythm.

A good issue in your voice has a recognisable pulse. Maybe you run a long build-up sentence and then drop a three-word fragment on top of it. Maybe you never do. Either way, the pattern repeats from issue to issue, and readers learn it without knowing they've learned it. That's the asset.

The same goes for openers. Some writers open every issue with a one-line claim. Others start with a small scene, or a question, or a number. The first six words of a paragraph say a lot about who's talking. And length matters more than people admit: an issue that normally lands around 500 words (the default newsletter length in ScriptGrain, with a range of 250 to 5,000) feels wrong at double that, even if every sentence is fine.

Then the small stuff. Contractions or none. How many commas you scatter per sentence. Whether you reach for dashes, semicolons, brackets. Whether you say "I" or "you" or "we". None of it is dramatic. All of it is noticed, because your readers see your punctuation every week.

You can't list all of this in a brief. Nobody can. You don't know half of your own habits until something measures them.

Where AI newsletters drift

Generic AI issues give themselves away fast. You've probably spotted them in other people's inboxes already.

The sentences are too even. Twenty words, twenty-two, eighteen, twenty-four. A human writer is lumpier than that; the AI smooths it out. Openers get formulaic, usually a sweeping statement about how fast something is changing. Transitions arrive on schedule. Every paragraph is roughly the same size, and every issue closes with a tidy little reflection.

Hedging is the other tell. A model told to sound "friendly and authoritative" tends to land on a safe middle: a few "might"s, a couple of "often"s, no position strong enough to be wrong. If you're a writer who states things plainly, that softening reads as someone else.

And then there's drift across issues. One draft might be close. The next wanders. By the fifth you've got a different narrator, and voice drift is exactly what readers register before they can say what changed. A tone description in a prompt doesn't fix it, because "warm but sharp" means something different every time a model reads it.

To be fair, a general-purpose model can write clean, competent prose. Nobody's disputing that. The issue is that clean and competent is the average, and you're not the average.

A worked example

Take an illustrative brief: an issue about why small teams should audit their own onboarding emails before buying new tools. Audience is existing subscribers, about 500 words, outline approved first.

Without a profile, the draft opens with a broad line about how onboarding has never mattered more. Paragraphs run even. Every point gets a hedge. The close gestures at "the journey ahead".

With the profile in play, several things change, and you can describe them without inventing numbers. The opener shifts to whatever this writer actually does: say, a blunt claim in the first sentence, followed by a fragment. Sentence length gets lumpier, matching the writer's own variance. Contractions come back if the writer uses them, and stay out if they don't. The confidence setting moves: declarative claims stay declarative instead of being padded with qualifiers. Paragraph openers pick up the writer's habitual words. The closing line shrinks to whatever sign-off pattern this person genuinely uses, which might be wry, or might be nothing at all.

The content is the same. The audit idea, the argument, the examples. What changes is who sounds like they're saying it.

What to feed the profile

Feed it issues you were happy with. That's the whole instruction, really.

A profile needs at least one sample of 50 words or more, but three or more pieces and around 3,000 words is the recommendation. For a newsletter, that's roughly six back issues at default length. Pick the ones you'd send again. Skip the rushed one, the one an editor rewrote, the one you wrote in someone else's register for a sponsor.

Mix matters too. If your newsletter has two modes, say a short commentary issue and a longer essay issue, include both, so the profile captures the range instead of one mood. If you'd like to see what gets measured before you commit, the free writing style analysis tool shows the countable part of your style in your browser, and the writing voice attributes reference lays out the 45 attributes and the eight layers they sit in: lexical, syntactic, tone and register, rhetorical, punctuation and format, function words, content patterns, and quirks and cadence.

Extraction is quick. The median measured on 2026-09-28 was 39 seconds. So the cost of trying it is small, and the cost of feeding it bad samples is a profile of the wrong voice.

Keeping issue forty consistent with issue one

Every draft gets scored from 0 to 1 against the profile. That's the voice match score. On-voice drafts typically land between 0.85 and 0.95. Partial matches sit between 0.5 and 0.8. Anything under 0.45 is clearly a different voice.

Before an issue goes out, read the number, then read the notes under it. The score alone tells you there's a gap. The per-feature deltas tell you where: draft value against profile value, feature by feature. Maybe your sentence-length variance came out flat. Maybe comma density is high, or the draft hedges where you don't. Those are fixable in a couple of edits, and you know which ones to make.

If the score's short of target, Polish revises the draft toward it (default target 0.9). And edits you save feed back in, so the profile keeps learning what you actually keep and what you cut. The method behind the scoring is published in the voice measurement framework, so you're not taking the number on faith.

A sensible routine is three checks, in order:

  1. Read the overall voice-match score.
  2. Scan the per-feature deltas for anything far from your profile value.
  3. Run the draft through the AI cliché checker for stock phrases the profile wouldn't catch.

That's it. Ten minutes, mostly reading.

Drafting from the assistant you already use

You don't have to leave ChatGPT or Claude. The profile reaches them through the MCP server and the REST API, and it's read live each time you draft. It's never a pasted prompt, which matters: a pasted block goes stale the moment your voice shifts, and it eats into your context window besides.

So you open ChatGPT or Claude, connect the profile, and ask for the issue as you normally would. The assistant pulls your measured voice as a constraint while it writes. The same connection works with Gemini, Cursor, Meta Muse and any MCP client, and both the MCP server and the API are on every plan, Free included. Details are on the API and MCP server page.

Both assistants are genuinely good at the parts around the voice: restructuring an argument, finding a sharper example, cutting a flabby paragraph. Keep using them for that. The profile just makes sure the draft they hand back sounds like the person whose name is at the bottom.

That's the arrangement. You write the ideas; the profile holds the voice steady. Issue forty, issue one, same narrator. Your readers get the same narrator every week, which is the point.

Other formats