Frame, then facts, then structure, then language. Each pass can invalidate everything below it, so reviewing in the wrong order means redoing work — and the bottom is the tempting place to start.
- Facts are where errors concentrate: specifics that appear rarely in training cannot be reliably recalled — the mechanism is on why AI makes things up.
Reviewing AI output in the wrong order wastes effort. Fix the frame before the sentences, because fixing sentences inside a wrong frame is work you will throw away.
Frame → facts → structure → language. Each pass can invalidate everything below it, so going upward means redoing work. Language is the tempting place to start, because it is the easiest to see.Reasoning, September 2026 — this order is the site’s own method, argued from dependency: a wrong frame makes the fact-check wasted, and wrong facts make the structure wasted. It is not measured, and no study comparing review orders is cited here.
Pass 1 · Frame
Before reading a single sentence for quality: is this answering the right question?
- Does it address what you asked, or what it inferred you meant?
- Did it accept a premise in your prompt it should have challenged? A loaded question produces a loaded answer — see how AI influences you back.
- Is anything important missing rather than wrong? Absence is hard to see, because nothing on the page points at it.
If the frame is wrong, stop. Do not fix anything else. Rewrite the brief and start fresh — reprompting inside the same conversation keeps the wrong frame in context, where it can anchor the next answer (if your tool’s memory is on, use a temporary chat; more on loops that degrade).
Pass 2 · Facts
Every specific. Names, dates, figures, citations, quotes, technical claims, legal references.
This is where errors concentrate, because facts appearing rarely in training cannot be reliably recalled: a formal result puts a floor under the error rate on facts that appear only once — the mechanism is on why AI makes things up.Kalai, Nachum, Vempala & Zhang, Why Language Models Hallucinate, 4 Sep 2025, §1.1, read at source 23 Sep 2026: the hallucination rate “after pretraining, should be at least the fraction of training facts that appear once.” A formal lower bound, not an empirical ranking of which kinds of detail fail most.
Check the claims that would be expensive to get wrong, not a sample of the easy ones.
Practical shortcut: anything confident and specific that you did not supply is the highest-risk sentence on the page. Start there.
Pass 3 · Structure
Now that the frame holds and the facts stand, does the argument work?
- Does each section earn its place, or is it there because the format suggested it?
- Is the strongest point buried? In this site’s experience, generated drafts tend to order by convention rather than by weight.
- Are there three examples where one would do? Padding is a common failure of generated structure, again in this site’s experience rather than by measurement.
Pass 4 · Language
Last, and quickly. By this point the frame is right, the facts hold and the structure works — language is the cheapest thing to fix and the least consequential to get wrong.
The one thing worth catching: confident phrasing on uncertain claims. Generated prose tends to sound as sure of a weak point as of a strong one, so hedging has to be added deliberately.Kalai, Nachum, Vempala & Zhang, Why Language Models Hallucinate, 4 Sep 2025, §1 and §1.2, read at source 23 Sep 2026: “Language models are known to produce overconfident, plausible falsehoods”, and “Bluffs are often overconfident and specific”.
When to stop
- Cap the passes. Pick the number before you start — see loops that improve and loops that degrade.
- Judge against the brief, not the previous version. Comparing to the last attempt pulls you toward the last attempt.
- If pass 1 fails twice, the task is wrong for this. Two failed frames is information about the task, not about the prompt.
Kalai, Why Language Models Hallucinate. Read at source 23 Sep 2026.