Making AI write like a person

AI writes a lot of our first drafts. Website copy, product pages, emails, documentation. Almost none of it ships the way it comes out, because raw model prose has a sound, and readers have learned to hear it. The moment your website sounds like a model, it reads as effort you did not spend, and customers discount everything it says.

We deal with this the same way we deal with every recurring AI problem: a written style guide that the agents load before touching customer-facing words. This guide walks through the actual rules in that file and why each exists. Steal any of it.

Know the tells

Model prose fails in patterns, and the patterns are consistent enough to list. Our file keeps a concrete ban list, and every item on it came from catching the pattern in our own drafts:

• 01The mirrored contrast: “It's not a template, it's a system.” Models reach for this shape constantly because it feels profound and costs nothing. One is fine. Five on a page is a signature.
• 02Rule-of-three rhythm everywhere: “Fast, flexible, and built to last.” Again, the shape is fine and the frequency is the tell.
• 03Em-dash pivots stitching every other sentence together.
• 04Winking meta-copy that narrates the site itself: “This is the part where we tell you about pricing.”
• 05Empty intensifiers: seamless, effortless, powerful, beautiful, all doing no work.
• 06Fake-precise numbers: “10x faster” and “100% satisfaction” with nothing behind them.

The fix is not asking the model to “sound more human.” That produces a model imitating a human imitating casual. The fix is banning the specific shapes, in writing, and rereading every visible sentence against the list before anything ships. Mechanical checks beat vibes: we literally grep our copy for em dashes before a page goes live.

Start from real writing, not from nothing

The strongest rule in our file is about where copy starts. Models asked to write from a blank page produce the average of everything they have read, which is exactly the generic voice you are trying to avoid. So our rule is that copy starts from live examples in the same register:

Match the rhythm of real studios and businesses in the same
register. Start from live examples, not invention. Read every
headline and CTA aloud once; if it sounds like a pitch deck or
a model being charming, rewrite it flatter.

In practice: before writing a restaurant website's about page, the agent reads three real restaurant about pages we respect and matches their plainness. Real businesses write shorter sentences than models do. They state prices without ceremony. They skip the part where the paragraph warms itself up. Reading them first calibrates all of that in a way no instruction can.

Write from a point of view

Voiceless copy is a tell all by itself. Somebody specific is supposed to be talking. On marketing surfaces we tell the agent exactly who: a marketing person who believes in the product and wants the sale, writing sentences a customer could repeat to a friend. Our file states it plainly:

Marketing pages are written from the point of view of a
marketing person who believes in the product and wants a sale,
in plain sentences a customer would say back to a friend. Say
the true thing directly instead of framing it against a
strawman.

The strawman line matters more than it looks. Models love to sell by contrast, describing a bad alternative and then positioning against it. Confident businesses mostly do not do this. They say what the thing is, what it costs, and what you get, and the absence of defensiveness is itself persuasive.

Separate voices for separate audiences

One more structural rule: different business lines get different registers, and mixing them is a bug. Our website work is productized, so its copy names deliverables and prices plainly. Our consulting work is scoped per engagement, so its copy talks about capability and outcomes and asks you to get in touch. The style guide bans crossing those streams, which stops the classic AI failure of a consulting page suddenly quoting package prices that do not exist, or a product page hedging like a lawyer.

Related: internal vocabulary stays internal. Planning words like “segments” and “arms” and whatever shorthand your team uses in meetings are positioning input, not headline text. The public page uses the words customers use. We keep that as a hard rule because models faithfully reproduce whatever vocabulary you feed them, and nobody outside your planning docs searches for your internal nouns.

Make it enforceable

None of this works as taste held in one person's head. It works because it is a file in the repository that every agent loads, with a mechanical pass at the end: reread every visible string, check for the banned shapes, check the numbers are real. When we catch a new tell, it gets a line in the file the same day, and then it stops appearing. That loop, catch, write down, never again, is the whole method. It is also, not coincidentally, how the rest of our agent harness works, which the harness guide covers in detail.

The membership includes the full playbook version of this with before-and-after rewrites, and the community where we workshop each other's copy. If you would rather we just write it, that is part of every website we build.