We are trying to make an AI write like a specific person rather than like an AI. Most of what we learn is unglamorous and specific, so we write it down here: the experiments that failed, the bugs that quietly changed the output, and the small decisions that turn out to matter.
It is rarely the vocabulary. Readers spot AI writing from structure, rhythm and a specific kind of empty confidence. Here are the tells we score every draft against, and how to remove them from your own writing.
One prompt behind every topic is why a post about fundraising reads exactly like a post about marathon training. What we found when we split ours into 48 separate engines, including the mistake that made every user in a niche sound identical.
We scanned five well-known X accounts, held everything constant except the voice profile, and generated a post for each. Two came back with the identical opening sentence. Here is what was actually wrong.