Skip to content
AI-Native PM
Field Notes

Field Note

How We Keep Our AI Drafts From Sounding Like AI

Every AI writing tool leans on the same handful of tics, and we were deleting them by hand from every draft. So we wrote the tics down once, as a plain set of rules any tool can follow, and put the file where anyone can take it.

· 5 min read

Read enough writing that came out of an AI tool and you start to see the same fingerprints: the em-dashes, the word "unlock" standing in for anything good, the openers that promise a hidden secret before landing on something ordinary, and the short sentences stacked three deep for a beat of drama. None of it is wrong on its own, but together it reads as machine-made, and once you can see it you cannot unsee it.

We hit this the moment we started drafting this site with AI in the loop. The same tics came back in draft after draft, and we found ourselves fixing the same handful by hand every time. At some point the obvious move was to stop correcting them one page at a time and write them down, so the tool could avoid them before we ever read the draft.

The fix for a repeated editing note is not more editing. It is a written rule the tool follows before it drafts.

So the rules grew the way scar tissue does, one note at a time. Every time one of us flagged something in a draft, we turned the flag into a plain, numbered rule and added it to a single file. Some rules are single words that had gone soft from overuse, each kept with a note on when the word is still the right one. Others are sentence patterns to avoid, written with a corrected example beside the broken one so the rule never stays abstract. The rest are habits of plainness: say the thing the way you would say it out loud, lead with what the reader needs, and never make anyone decode a sentence to reach an ordinary point.

The file paid for itself the first time we pointed a fresh tool at it. Instead of re-explaining our preferences from scratch in every new chat or repo, we hand over the one file and get drafts that already sound like us. When our taste changes, we change the rule in one place, and every tool that reads the file moves with it.

Most of that file is specific to this site: our voice, our course structure, the research we lean on. But the core of it, the banned words and the sentence tics and the rules for plain writing, is not ours alone. Every one of these tells shows up in everyone's AI drafts. So we pulled the portable rules into a second file with nothing site-specific left in it, and put it in a public repo. If you write anything with an AI tool, you can take it, drop it into Claude, ChatGPT, Copilot, or your coding setup, and bend it to your own voice.

The file lives here: content-rules-generic.md on GitHub. Take it, use it, and if you catch a tell we missed, tell us so we can add the rule.

Where this goes next

Writing the rules down is one case of a habit that runs through this whole build: give the AI standing instructions instead of correcting the same thing forever, which is the subject of Why We Told the AI to Disagree With Us. And the reason those plain-writing rules earn their place is the reader on the other side of the page, whose attention is finite and easily spent; that is the ground our essay The Human Factors covers.

Sources