We Ran One Prompt on 3 of Our Own Paragraphs. It Never Wrote a Word. | Edition 291
Edition 291 — One interview-style prompt turns any chatbot into a line editor, not a writer. We ran it on ourselves first.
Every "write me a first draft" prompt makes the same trade: you get words fast, and you get someone else's voice. The fix is not a better generation prompt. It's not generating at all.
We built an interview-style prompt that never writes a single new sentence. It only edits the paragraph you already wrote — cutting what's dead weight, tightening what's loose, and flagging any claim you can't actually defend. Then we ran it, for real, on three paragraphs nobody wrote to look good for this newsletter.
Why editing beats generating
Ask a model to write your paragraph and it has to invent a voice — and it reaches for the same one every time: smoother, vaguer, and instantly recognizable as AI. Ask it to edit your paragraph instead, and it has nothing to invent. It just has your sentences, and a job: cut what doesn't earn its place, tighten what's loose, and flag what you can't back up.
The mechanism that makes this work isn't the editing instruction — it's the interview. Before it touches a word, the prompt asks who's reading, what's actually bothering you about the draft, which phrase is off-limits, and which claims you're not sure hold up. Skip that step and any model defaults to smoothing everything into the same tone. Answer it, and the model has specific, defensible reasons for every cut.
What happened when we actually ran it
We tested the prompt live, three separate times, on three unscripted paragraphs — a startup update, a career-change post, and a confession about using AI to help write this very newsletter. Every run followed the same shape: the model asked its five questions one at a time and waited for real answers, then handed back a tightened paragraph, a list of what it cut and why, and a flagged list of claims that needed a source.
It preserved the exact phrases we told it to protect — a shaky-handed table-saw confession, a three-sleepless-weekends line — while cutting the cliches and softened stats sitting around them. On the paragraph with a made-up statistic, it didn't invent a source or quietly delete the number. It said, in plain language, what proof would be needed before the claim could stand.
| Test paragraph | Voice line it protected | Claim it flagged |
|---|---|---|
| Startup onboarding update | "I built this thing over three sleepless weekends... I'd do it again in a heartbeat." | 40% activation lift — still being validated internally |
| Career-change post | "My hands still shake a little every time I turn on the table saw." | 80% career-changer regret stat — admitted invented on the spot |
| AI-assisted newsletter confession | "Have the AI ask me harder questions about what I was trying to say." | 25% open-rate drop — never actually pulled from analytics |
None of this required a smarter model. We ran it on a mainstream, current chatbot, and it worked because the prompt does the hard part before generation ever starts: it asks, it waits, and it only touches what's actually broken.
The prompt is below, exactly as we ran it. Interview it with your own paragraph and see what it leaves alone.