Notes · No. 09

Why your Copilot still sounds like a robot.

Leigh Daniel · August 2026 · 5 min

Your company rolled out an assistant. Copilot in Word, maybe, or Gemini in Docs, or a chat window with your name on the account. Somewhere in its settings is a promise: it can learn to write like you. You gave it a sample. The first paragraph came back wearing one of your phrases like a name tag. By the third paragraph it was a robot again.

You didn't do anything wrong, and neither, exactly, did the tool. It did what the setting does. The setting just isn't what you thought it was.

What the setting actually learned

When one of these tools "matches your style," here is the whole mechanism: it reads your sample and produces a few sentences of guidance about it. Professional but warm. Direct. Uses short paragraphs. Adjectives. Then it writes with those adjectives taped above the desk.

That's a real feature and it does real work. If your old drafts came back sounding like a press release, the setting will get you to "sounding like a person." The problem is the gap between sounding like a person and sounding like you, because you are not three adjectives.

What your voice actually is

Your voice is structural. It's the shape of your sentence lengths: a long one that sets something up, a short one that lands it. It's where you put the ask (first line, if you're the Director we write for). It's the moves you repeat without noticing, the analogies you reach for, and the list of things you never do: never "hope this finds you well," never three hedges in a row, never a triad where one item would land harder.

None of that fits in three adjectives, so none of it survives the setting. Tone is a costume. Voice is a build.

Why it drifts back

Even the parts the setting caught don't hold. The guidance sits at the top of the draft like a sticky note, and the longer the writing runs, the more the model's own habits take over: the rhythm flattens, the hedges creep in, the punctuation starts doing that thing. Every one of these tools has a default voice, and the default always comes back. You've watched it happen by paragraph three. That's not a bug in your instance. It's what a note taped above the desk does when the desk gets busy.

The setting learned three adjectives. Your reader knows you as a rhythm, a set of moves, and a list of things you'd never say.

The fix isn't a new tool

Here's the part the tool comparisons miss: you don't need to replace Copilot. You need to arrive with your voice written down.

That's the whole idea of a voice profile: a document that captures the structure, not the vibe. Your rhythm, measured. Your signature moves, named, with your own sentences as the evidence. Your never-list, itemized. It's specific enough that a model can actually follow it, and it's a file you own, which means it goes wherever you write: into Copilot, into Gemini, into Claude or ChatGPT, into whatever your company rolls out next year.

Give the model your patterns before it writes, and check what comes back against your never-list after. That's the entire method. Here's what a full profile looks like, including the before and after on a status update, so you can judge the difference yourself instead of taking our word for it.

Your Copilot isn't broken. It's just writing for the average of everyone, because nobody handed it you. Hand it you.

The profile in that specimen was built with the Voice Kit: a worksheet, a builder that runs in your account, and the filled instruction you leave with. One price, once. The profile is a file you keep, exportable from day one, and it works in every tool named above. If you'd rather start free, the Slop Check will show you which tells your current drafts trip.