A short writing sample gives AI a confident, wrong voice

Why a small sample fails quietly

A real first attempt at a voice specification was built from roughly two hundred words of email. What came back read as salesy and hyped, an exclamation-mark voice that wasn't the author at all. A small sample doesn't give you less of the picture. It gives you a confident, incorrect one, and the failure mode is a model guessing confidently from too little evidence rather than guessing badly. A bad guess that looks uncertain gets checked. A bad guess that looks certain gets shipped.

What to collect, and how much

Gather tens of thousands of words of working writing: emails, messages, transcripts of you explaining something to somebody else. Not published writing, because published writing has already been sanded toward a generic middle, and that middle is exactly the register you're trying to avoid.

One documented voice corpus ran to 1,544 messages and 49,068 words, pulled from a two-year export of real messages. It's worth naming plainly that this specific corpus was built for a colleague, not for the author himself; borrowing someone else's properly measured result and presenting it as proof of your own method would be exactly the kind of confident, wrong answer this whole exercise exists to prevent.

What done looks like

A corpus of real, unpublished writing exists, in the tens of thousands of words, pulled from genuine working correspondence rather than anything already polished for an audience. That corpus becomes the raw material the next step (locking the voice specification) measures against, rather than guessing at from a handful of paragraphs.