By promptcraftasia.com Team
PROMPT ENGINEERING

AI Critique Prompts: Get Honest Feedback on Work

Twenty new prompts that make AI read your drafts the way a sceptical client, assessor, recruiter or customer would, and tell you what's wrong before anyone else does.

·11 min read

Quick Answer

A good AI critique prompt names a specific, sceptical reader, asks for problems ranked by how much they hurt your goal, and tells the AI not to rewrite your work or praise it for the sake of it. It also gives the AI a way out: if the work is fine, say so and stop.

The best AI critique prompts tell the model exactly who is reading your work and what that person is looking for, then ask it to find problems rather than fix them. "Read this proposal as the client's finance director, who thinks it's too expensive" gets you a real review. "Can you give me feedback on this?" mostly gets you a compliment and a light polish.

That gap is the reason for this week's prompts. Most people already use AI to write things. Far fewer use it to check their own work before a client, a teacher, a grant assessor or a customer sees it. That's a shame, because a second reader that never gets tired and never worries about your feelings is exactly what most drafts need.

The catch is that AI models don't naturally behave like that second reader. They lean towards agreeing with you. So these twenty prompts, two in each of the ten categories, are written to push against that habit.

What Are the 20 New Prompts For?

They fall into four groups, by the kind of reader they stand in for.

The sceptic. Find the Weakest Argument in Your Draft, List the Questions Your Boss Will Ask About a Plan, Test Your Salary Case Before the Conversation, and Check Whether Your Bug Fix Hits the Root Cause. These look for the claim or assumption that won't survive a hard question.

The person on the other side. Poke Holes in a Proposal as the Client Would, Find Why a Cold Email Would Be Ignored, Find the Questions Your Reply Leaves Unanswered, Review a Product Listing as a Doubtful Buyer, Find the Weak Spots in Your Interview Answers, and Predict Why a Post Might Flop Before You Publish. These read your work as the person you want something from.

The assessor with a checklist. Get Rubric Feedback on Your Essay Without a Rewrite, Critique a Grant Application Like an Assessor, and Critique a Generated Image Against Your Brief. These score your work against criteria you paste in.

The newcomer and the stress test. Test What a Skimming Reader Takes Away, Review Your README as a New Developer Would, Critique a Lesson Plan as a Struggling Student, Check How a Post Reads Across SEA Markets, Find Why Your Image Prompt Gives Mixed Results, Challenge Your Weekly Plan for Overcommitment, and Get Feedback on Any Work Without the Flattery. These find the places where someone new, rushed or from a different market gets lost.

The writing prompts sit in Writing. The grant and listing prompts are in SEA Business, since they deal with local schemes and marketplaces.

Why Does AI Go Easy on Your Work?

Because it was partly trained to please people, and people like being agreed with. Researchers call this sycophancy, and it's well documented.

In a study published at ICLR 2024, Mrinank Sharma and colleagues found that five state-of-the-art AI assistants "consistently exhibit sycophancy" across four different writing tasks. When they looked at the human preference data used to train such models, they found that a response matching the user's views was more likely to be preferred. Both humans and the preference models trained on their judgments sometimes picked a convincing, agreeable answer over a correct one.

A 2025 benchmark from Stanford researchers goes further. ELEPHANT, by Myra Cheng and colleagues, measures "social sycophancy," which is the habit of protecting the user's self-image. Across 11 models, they found that LLMs preserved the user's face 45 percentage points more than humans did in general advice questions and in posts describing clear wrongdoing by the user. When given both sides of the same conflict, the models told both parties they weren't in the wrong in 48% of cases, depending on which side the user took.

For feedback, that's a real problem. If you say "I think this proposal is strong, can you check it?", you've told the model what answer you'd like. It may well give it to you.

The good news is that how you frame the request changes what you get. The rest of this guide is about framing.

How Do You Get a Critic Instead of a Cheerleader?

Give the AI a reader to play, and make that reader hard to please. Every prompt this week does this in its first line.

Compare these two openings:

The second one gives the model a point of view, a set of priorities and, implicitly, permission to object. It also changes what "good" means. A proposal can be well written and still fail the operations manager's test, and that's the test that decides whether you get paid.

Three other habits make the critique sharper:

This builds on role prompting, which we cover in more depth in our role prompting guide. The difference here is that the role isn't an expert who helps you. It's a reader who has to be convinced.

Why Ask for Problems Instead of a Rewrite?

Because a rewrite hides the critique, and it usually costs you your own voice.

When you ask an AI to "improve" a draft, it changes everything at once. You get a smoother version, but you can't easily see what was wrong with yours or whether the changes were needed. And if the text goes out under your name, a client or manager who knows how you write will notice.

So most of this week's prompts include a line like "Do not rewrite the email. Point to the lines and explain the problem." That gives you three advantages:

The two image prompts are the exception. For Critique a Generated Image Against Your Brief and Find Why Your Image Prompt Gives Mixed Results, a revised prompt is the useful output, so they ask for one. Even then, they ask the AI to explain each change in a line, so you can see the reasoning.

When Should You Give the AI the Marking Criteria?

Whenever someone else will judge your work against a written standard. That covers rubrics, grant criteria, a client brief, a job description or a design spec.

Without the criteria, the AI judges your work against its own idea of "good," which may have nothing to do with how you'll be assessed. With them, it can do something much more useful: go criterion by criterion and tell you which ones you clearly meet, partly meet or don't address at all.

Two of this week's prompts are built around this. Get Rubric Feedback on Your Essay Without a Rewrite asks the AI to place the essay in a band for each criterion and quote the evidence. Critique a Grant Application Like an Assessor does the same with a scheme's published criteria, and adds checks for vague outcomes and claims an assessor would want evidence for.

The grant prompt also says "Use only the criteria I pasted. Do not add rules from memory, because schemes change often." That instruction matters. A model's training data may describe last year's version of a scheme, with different caps, eligibility or deadlines. Paste the current criteria from the agency's own website, and check that page again on the day you submit.

If you want to go a step further and check the AI's own claims, our guide to verifying AI output covers that side of the problem.

What Should You Do When You Disagree With the AI?

Push back, but watch what happens next. This is where sycophancy causes the most trouble.

If you reply "I don't think that's a problem," many models will quickly agree and drop the point, whether or not you're right. The ELEPHANT finding that models told both sides of a conflict they were in the right, in 48% of cases, is a warning here. The model's view can shift with yours.

The prompt Get Feedback on Any Work Without the Flattery tries to guard against this with a rule set at the start: "If I push back on a point later, do not drop it just to agree with me. Change your view only if I give you a reason." It's not a guarantee. But it sets an expectation for the rest of the conversation, and it gives you a way to tell a real change of mind from a polite retreat. If the model changes its view, ask it what reason you gave that changed it.

A second safeguard is to argue with the point, not the feeling. "That section is fine because the client asked for exactly that format in their brief" is a reason the model can weigh. "I like that section" isn't.

And in the end, the critique is advice. You know the client, the class or the market. Use the AI to see problems you'd miss, then make the call yourself.

How Do Critique Prompts Help Teams in Asia?

Because AI is already everywhere at work here, and much of that work crosses languages and markets.

IMDA's Singapore Digital Economy Report 2026 found that 85.7% of surveyed workers used AI in 2026, up from 77.6% the year before. AI adoption among SMEs rose from 14.5% in 2024 to 23.4% in 2025. Yet only 36.7% of workers had attended AI-related training in the past 12 months. Plenty of people are using these tools without being shown how to use them well, and asking for an honest review is one of the easiest skills to pick up.

Two prompts this week deal directly with the region:

The cold email and proposal prompts help too, in a quieter way. Many SME owners in the region write business English as a second or third language. A critique prompt that points to the exact line that sounds vague or pushy is often more useful than a full rewrite in someone else's style.

Which Prompts Should You Run First?

Start with whatever you're about to send. Critique is most useful just before something leaves your hands.

Once you're used to the pattern, Get Feedback on Any Work Without the Flattery works as a general-purpose reviewer for almost anything. All twenty prompts are written for ChatGPT, Claude and Gemini alike.

What Else Do People Ask?

Why does AI always say my work is good?

Partly because models are trained on human preferences, and people tend to prefer answers that agree with them. Researchers call this sycophancy. You can reduce it by naming a sceptical reader, asking for problems ranked by impact, and telling the AI it may say the work is fine if it genuinely is.

Can I use AI to get feedback on a school essay?

Often yes, but check your school's AI policy first. Many schools allow feedback on your own draft but not AI-written text. Paste the marking rubric, ask the AI to place your essay in a band for each criterion with quoted evidence, and tell it not to rewrite any sentences.

Should I tell the AI who will read my work?

Yes. It's the single biggest improvement you can make. A named reader with clear priorities, such as a client's finance director who worries about cost, gives the AI a standard to judge against. Without one, it judges against a generic idea of good writing.

Is it safe to paste client documents into an AI tool?

Check your company's policy and your client agreement first. Remove names, contact details and other personal data you don't need for the review. For sensitive work, use a business or enterprise plan your organisation has approved, and read how that plan handles your data.

Do these critique prompts work in ChatGPT, Claude and Gemini?

Yes. They're plain instructions with no tool-specific features, so they work in any major chat assistant. The image critique prompt needs a model that can read an uploaded image. Results will vary between models, so judge each point on its merits.

Sources: Sharma, Tong and colleagues, Towards Understanding Sycophancy in Language Models, ICLR 2024, for sycophancy across five AI assistants and its link to human preference data. Cheng, Yu and colleagues, ELEPHANT: Measuring and understanding social sycophancy in LLMs, 2025, for the 45 percentage point and 48% findings across 11 models. IMDA, Singapore Digital Economy Report 2026, for worker and SME AI use and training.