The fastest way to review AI content is to stop reading it like a proof-reader. If you go top to bottom, fixing each sentence as you meet it, you'll spend an hour polishing a paragraph you were always going to cut, and you'll still miss the wrong fact in the middle. Review big to small instead: structure first, facts second, wording last. Fix as you go, in the document, not by re-prompting.
AI drafts invite a particular kind of over-reviewing. The prose is fluent and confident, so your eye slides along at sentence level, tidying commas, while the actual risks, a made-up statistic, a claim that doesn't hold, a section in the wrong order, sit unexamined behind the polish. Fluency hides the errors that matter. A structured pass beats a linear one.
Pass one: is the shape right?
Before you touch a single word, step back and ask whether the whole thing is built correctly. Does it make the right argument? Is it aimed at the right reader? Is anything important missing, and is anything here that shouldn't be? Is the order logical?
This pass takes a couple of minutes and saves the most time, because structural problems make everything downstream pointless. There's no sense perfecting a sentence in a section you're about to delete. If the shape is wrong, that's the moment to go back to the model, because a structural fix is a re-prompt job, not an editing one. If the shape is right, you've just protected yourself from an hour of wasted polishing. Get the skeleton right before you worry about the skin.
Pass two: are the facts true?
This is the pass people skip, and it's the one that will actually embarrass you. AI content is at its weakest on specifics: names, numbers, dates, titles, quotes, anything that has to correspond to a real thing in the world. It will state all of them with the same smooth confidence it uses for the safe general points, which is exactly what makes them dangerous.
So go through deliberately and mark every concrete claim. Every figure, every name, every "studies show," every specific assertion about a person or company. Then verify the ones that matter, or cut them. Treat an unsourced specific as guilty until proven innocent. This is slower than the other passes and worth every second, because a single wrong fact in something you send to a client or publish undoes all the time the AI saved you and then some. If it's going to a third party, nothing concrete goes in that you can't stand behind.
Pass three: fix in place, then polish
Only now, with the structure sound and the facts checked, do you touch the wording. And here the how matters as much as the what. Fix errors directly in the document rather than describing them back to the model. Re-prompting to correct a near-finished draft is a bad bet: it can rewrite the parts you were happy with, and you'll have to re-review the whole thing to catch what changed.
Editing in place keeps you in control and keeps the review contained. Change the wrong line, leave the rest, move on. The friction here is real, especially with HTML documents where the text is tangled in markup and changing a heading means hunting through tags. That friction is what tempts people back into re-prompting. Removing it is the point of htmlEdit: you edit the rendered document visually, clicking and typing on the page, so a fix costs seconds and never risks the rest of the draft. I go deeper on this in Stop Re-Prompting.
Give it a time budget
One more thing, because review expands to fill whatever time you give it. Decide up front how long this is worth. A quick internal note deserves ten minutes. A client-facing document deserves more, but still a number, not "until it's perfect." Setting the budget in advance stops the quiet drift where a fluent draft lures you into endless small tweaks that make no real difference.
The whole method is really one idea: spend your attention where the risk is. Structure and facts are where AI content fails and where errors cost you. Wording is where it's already strong. Review in that order, fix in place, cap the time, and you'll get through it in a fraction of an afternoon with a better result than a linear read would ever give you.