The Three Places Generic AI Output Actually Gets Fixed

If your AI-assisted work has started to read like everyone else's AI-assisted work, the fix isn't a sharper prompt template or a different model. Four unrelated stories from this year land on the same three checkpoints in your process — and most people are only guarding one of them.

Four stories, one pattern

Figma's CEO Dylan Field told Stratechery that AI output "draws from the middle of the distribution," and that teams get viscerally attached to whatever concept it hands back first — which is exactly why the first draft becomes the ceiling instead of the start of the work.

The same month, two Hacker News posts with a combined 1,374 points named the two seams around that middle draft: a vague brief going in, and raw, unedited text going out. Neither post blamed the model.

MIT Sloan found the input problem has a price tag: AI financial advice built on a structured prompt — age, income, goals, timeline — produced simulated outcomes worth roughly $50,000 more by age 60 than advice built on a vague one. Same tool, same access. The gap was entirely in who already knew how to ask a complete question.

And in September, Simon Willison endorsed a rule blunt enough to hold: write the draft yourself, use AI only to flag what's broken, never to supply the fix.

None of these four are about model quality. They're about three checkpoints — what goes into the prompt, whether you stop at the first thing the model hands back, and what leaves your hands unedited.

The three checkpoints, Monday morning

Before: put in the specific fact, not just the complete data. A prompt with every number but the wrong goal produces confident, well-organized, wrong advice. Write down the two or three facts a competent human in that role would ask you for first, and check they're actually in the prompt before you send it.

During: don't stop at the first thing back. The first draft is competent by design — it's the average of a huge distribution, which is exactly why it reads like it could work. Generate a second version with a deliberately different angle before you decide the first one is good enough.

After: the last sentence has to be yours. Ask AI to flag what's unclear, too long, or hedging — never to hand you the fixed version. Fixing it in your own words is the whole point; pasting in its rewrite undoes the first two checkpoints.

Priya: guarding all three

Priya runs marketing at a 40-person SaaS company and uses AI to draft the monthly customer newsletter. Her brief now opens with three lines before anything else: the segment reading it, the one action she wants them to take, and the specific objection last month's low-performing send got in replies. She generates two versions from that brief — one following her usual structure, one told to ignore it — and picks between them instead of editing the first one that lands. Before anything ships, she rewrites the opening line herself; it's the one sentence she never lets the draft keep as written, because it's the one a repeat reader would notice first.

Tomas: skipping the last one

Tomas manages client updates at a 90-person consulting firm and had all three checkpoints working — a tight brief, a second draft compared against the first — until deadline pressure made him paste the AI's polished version straight into a client email without rewriting the closing line himself. That line described a project risk as "being actively monitored," a phrase the model supplied, not a status Tomas had actually verified. The client noticed something was off before Tomas did: the sentence was true-sounding in a way none of his other updates were. He now reads every closing line out loud before sending, specifically checking that it's a claim he could defend if asked.

The one-sentence version

A better model given a vague brief, stopped at the first draft, and shipped unedited will still produce a generic, unverified, unmistakably average result — the checkpoint was never the model's to guard.