Ask an AI to review something you made — a cover letter, a business plan, a short story — and it will usually tell you it’s great, then offer a few gentle suggestions buried under compliments. This is the default behavior, and it’s nearly useless if your actual goal is to improve the thing.
The problem isn’t the model. It’s that “give me feedback” is underspecified in exactly the way that produces the blandest possible response. The model doesn’t know whether you want validation or demolition. It doesn’t know what you’re planning to do with the feedback. It doesn’t know which parts you’re already uncertain about. So it hedges by being nice, covers its bases with a couple of surface-level notes, and calls it done.
You can fix this — not with magic prompt templates, but with a slightly more deliberate setup each time you ask.
Start with the role and the stakes
Before you paste your work in, tell the model who it’s supposed to be and what would happen if your thing failed. Not vague stuff like “be a harsh critic” — that just gets you a slightly less polite version of the same feedback. Something more like:
“You’re a hiring manager who reads 80 cover letters a week. You’re skeptical. Read this one and tell me the first moment you’d stop believing it or want to move on. Be direct — I’m not looking for encouragement.”
That framing does two things. It gives the model a specific perspective to inhabit — not just “critic” but a particular type of critic with a particular agenda. And it signals clearly that politeness is not what you’re optimizing for. The model picks up on that second signal more than most people expect.
Tell it what you’re already worried about
This is the step people skip, and it’s the one that makes feedback most useful. If you have a nagging feeling that your third paragraph is weak, or that your pricing section might confuse people, say so explicitly:
“I’m worried the opening is too generic and the pricing section is unclear. Start there before anything else.”
Two reasons this works. Models are genuinely good at responding to a named problem — they’ll engage with it more specifically than they would if they had to hunt for weaknesses on their own. And naming a flaw yourself short-circuits the politeness default: if you’ve already flagged it, the model isn’t delivering bad news anymore. It’s just answering a direct question.
Ask for a specific failure mode, not general impressions
“What could be better?” produces a list. Usually a bland one. More useful questions:
- “What’s the single weakest sentence in this, and why?”
- “Where does the logic break down — where would a skeptical reader stop trusting me?”
- “If you had to cut 30% of this and keep the core argument intact, what goes?”
- “What question does this raise that it doesn’t answer?”
These work because they force a concrete answer. The model can’t say “the tone could be stronger” when you’ve asked it to find the single weakest sentence. It has to commit to something specific, which gives you something real to work with — even when you disagree with the answer.
Run the feedback in rounds, not all at once
Dumping everything into one prompt and asking for comprehensive feedback is almost always the wrong move. You get a long list of points with roughly equal weight, and it’s hard to know what actually matters versus what the model included to seem thorough.
Better: ask about one thing at a time. Start with structure or argument. When you’ve dealt with that (or decided to ignore it), ask about clarity. Then tone. Then the specific thing you were already worried about.
Each round stays short and actionable. The less obvious benefit: by the third or fourth exchange, the model has context on what you’ve already addressed and what you’re still working on, which makes later feedback more precise. It’s not starting from scratch each time.
This is the actual feedback loop. Not a one-shot review, but a back-and-forth where you’re actively deciding what to do with each piece of feedback before asking the next question.
Push back when something feels wrong
AI feedback isn’t automatically correct just because it sounds confident. Models will sometimes flag things as problems that are actually deliberate choices — a short punchy sentence that reads as “incomplete,” an unconventional structure that reads as “disorganized.” When that happens, say so:
“The short paragraph you flagged was intentional — I wanted a pause there. Does the issue actually affect the reader, or is it just an unusual choice?”
The model’s response usually clarifies whether it was giving you a real structural note or just pattern-matching against what “normal” looks like. Sometimes it’ll concede the point. Sometimes it’ll explain why the concern stands even with that context. Either answer is more useful than silently accepting or silently ignoring the feedback.
What this process won’t catch
AI feedback is genuinely useful for surface-level clarity, logical consistency, and structure. It’s much weaker on whether an argument is actually true, whether a joke lands with a real audience, or whether a business idea has legs. The model can tell you your pricing section is confusing — it can’t tell you whether your pricing is right. And it has no real stake in whether your thing succeeds, which means it won’t push back the way a colleague or a customer would.
Use it for the stuff it’s good at. Don’t let a clean round of AI feedback substitute for one real human reading the thing before it goes out.
Take one thing you’ve made recently that you’re not quite satisfied with. Instead of asking “what do you think of this,” ask “where would a skeptical reader stop trusting me?” Then disagree with at least one thing it says. That’s the posture that makes this useful.
