x
Paste this at the end of any long job, after the output has been produced:
"Now list what you were unsure about, flag anything you couldn't verify, and say which parts a human needs to check before this goes anywhere."
That's the whole tool. Fifteen seconds to type, and it changes what you're reviewing from a wall of confident prose into a document with a map.
The reason to run it after the work rather than inside the original brief is that a model asked to be uncertain up front tends to hedge everything, which is useless. Asked afterwards, about work it has already committed to, it produces something much closer to an honest inventory.
Because tone carries no information about accuracy, and humans read it as though it does.
This is the failure we've called a liability with good manners: a fluent paragraph built on thin evidence looks exactly like a fluent paragraph built on solid evidence. Same rhythm, same vocabulary, same air of having checked. There is no stylistic tell, which means you cannot catch it by reading carefully — reading carefully is precisely the skill it defeats.
So the fix can't be attention. It has to be structure. Asking the model to separate what it knows from what it assembled gives you something reading alone can never produce: a shortlist. Instead of reviewing 900 words at uniform suspicion, you review the four lines it flagged at high suspicion and the rest at normal.
That is a genuinely different activity, and it's faster. Most people find the flagged lines are where the errors actually live — not always, but often enough that starting there beats starting at the top.
It also changes what review feels like, which matters more than it sounds. Reviewing a long document with no map is an exercise in vigilance, and vigilance decays across a page in a way nobody can train out of themselves. Reviewing four flagged claims is a task with an end. The second one gets finished on a Friday afternoon; the first one gets skimmed.
A caveat worth stating plainly, because this tool gets oversold: the model's self-assessment is not reliable. It will miss things it was confidently wrong about, which is by definition the category you most need. Treat the flags as a priority ordering, never as a clean bill of health for everything unflagged. It tells you where to look first. It does not tell you where you needn't look.
Three moves, and the third one is the one that connects this to the rest of the week.
Check the flagged lines first. Names, dates, figures, citations, anything load-bearing. If a flag says "I couldn't verify this statistic", that is a research task, not a formatting note — and it's the one most likely to end up in front of a client.
Keep the note with the draft. Don't delete it once you've reviewed. Paste it at the bottom of the working file, or into the same folder. Six weeks later, when someone asks where a number came from, the uncertainty note is the fastest route to the answer — the same argument as a decision log, applied at the level of a single document.
Use it to say what AI actually did. This is where the tool earns its place in a disclosure week. When you tell a client "AI drafted this and I reviewed it", the uncertainty note is the evidence that the review was real and directed. It converts a claim into a record. If you've written the three-sentence disclosure, sentence two promises a human check — this is the artefact that shows one happened.
Three places, from practice.
Anything with a number in it. Reports, forecasts, comparisons. Numbers are where confident-sounding wrongness costs most, and they're the easiest thing for a model to produce plausibly.
Anything you'll be quoted on. Client-facing summaries, proposals, anything that becomes someone else's input.
Anything you didn't write the brief for. Work handed to you, work resumed after a week away, work produced from a voice note — cases where you can't remember what you asked for and therefore can't spot what's missing.
The place it earns least is short, low-stakes copy. A social caption doesn't need an uncertainty audit; it needs an editor. For that, make it cut rather than praise is the better tool, and running both is usually one step too many.
The underlying habit is neither "trust it" nor "don't trust it". Both are positions rather than processes. The habit is: make it show you where to look, then look there. That's the same instinct as asking a model to show its working — one asks how it got there, this one asks where it's least sure it got there correctly.
Try it on your next long draft. Fifteen seconds in, a shorter review out.
For the prompt set this line belongs to — free, 5-min read: 7 AI Prompts That Run a One-Person Business.
Related Decodes
AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.
TL;DR: AI writes with identical confidence whether it is certain or guessing, so careful reading cannot catch the difference — the fix has to be structural. At the end of any long job, add one line: list what you were unsure about, flag anything you couldn't verify, and say which parts need a human check. Run it after the work, not in the original brief, or the model hedges everything. The output is a shortlist, which turns a uniform-suspicion review into a prioritised one, and the flagged lines are where errors usually live. Treat it as a priority ordering, never a clean bill of health — a model misses what it was confidently wrong about. Then check the flags first, keep the note with the draft as a record, and use it as evidence that the human review you disclosed actually happened.
