Have you ever asked an AI "what do you think of my idea?" only to get showered with praise, then discover later the idea was riddled with holes? The problem is not that the AI is dumb; it is that you asked the wrong way. This article teaches a technique that works across ChatGPT, Claude, and Gemini: "rubric prompting," which turns AI from a people-pleaser into an objective critic, complete with a ready-to-use template and five worked examples.
What is rubric prompting?
Rubric prompting means this: instead of asking the AI for vague "feedback," you require it to score each item objectively against criteria you define in advance, with a reason for every score. A complete rubric has three parts: a score and scale, the criteria, and descriptors for each level. This forces the AI to weigh each dimension rather than reflexively praise you.
Why does AI default to flattering you?
After training, large language models are inherently inclined to please the user. When you ask an open-ended "what do you think?", their default mode is to go along with you and respond positively. This is not deliberate deception; absent a clear standard, they have nothing to measure against, so they default to praise. To break the pattern, you must hand them a measurable "ruler."
A real example: from glowing praise to 8 out of 100
One founder tested the same business idea two ways. Asked open-endedly, the AI praised it as viable. Switched to a rubric prompt requiring a 1-to-100 score against explicit criteria, the same AI gave it just 8 out of 100, itemising fatal flaws: a tiny market, fierce competition, and high costs. Same idea, same AI, only the question changed, and it saved the founder months of time and a great deal of money.
The three parts of a rubric prompt
An effective rubric has three components. First, a score and scale (for example 1 to 5, or weighted to total 100). Second, the criteria (the dimensions you want measured, such as feasibility, market demand, competition, cost, and risk). Third, descriptors for each level (spelling out what "excellent," "adequate," and "poor" look like). The more specific the criteria, the more reliable the output.
A full prompt template (copy and use)
You can apply this framework directly: "Using a scale of 1 to 100, objectively evaluate my [idea / copy / plan] against these five criteria: (1) market demand, (2) feasibility, (3) level of competition, (4) cost and risk, (5) uniqueness. Score each separately (20 points each) with reasons, then total them. Score from the perspective of a strict investor; do not inflate the score to encourage me." One instruction turns praise into scrutiny.
Five worked examples (copy directly)
Below are five rubric prompts covering different scenarios. Replace the bracketed content and use them as-is:
- Evaluate a business idea: "As a strict venture investor, score this business concept from 1 to 100 on: (1) market size, (2) authenticity of the pain point, (3) profit model, (4) barriers to entry, (5) competitive advantage. 20 points each, state why points were deducted, and name the single most fatal risk."
- Critique sales copy: "By the standards of a senior copy director, score this copy from 1 to 10 on appeal, clarity, persuasiveness, and call-to-action strength. Give each a score with specific rewrite suggestions; no pleasantries."
- Compare multiple options: "Here are three options. Using the same criteria (impact, cost, risk, speed to launch, 25 points each), score and rank each, then explain why the winner came first."
- Review a job application or bid: "As a hiring manager, score this resume from 1 to 100 on relevant experience, quantified results, clarity of expression, and fit for the role. Identify the three weakest areas and how to fix them."
- Inspect code or a workflow: "By the standards of a senior engineer, score this workflow from 1 to 10 on readability, efficiency, security, and maintainability, with reasons for each, and list the highest-priority issue to fix first."
Advanced: make the AI score itself (self-scoring)
At the end of a complex prompt, add one line: "When done, score your own answer from 1 to 10 on clarity, usefulness, and accuracy, and point out how it could be improved." This "reflective prompting" forces the AI to inspect its own output, find weaknesses, and revise, which has been shown to noticeably improve answer quality.
Two more ways to stop AI from giving lazy answers
Beyond scoring, two lesser-known but effective tricks help. First, ask "did you actually search for this?" to force it to genuinely use its built-in web search rather than guess from memory. Second, instruct it to "calculate this, do not estimate" to make it run real computation via code. These commands move the AI from "guessing answers" to "producing evidence."
Best use cases and common mistakes
Rubric prompting is most powerful when evaluating business ideas or investments, critiquing copy and proposals, comparing options, and conducting risk reviews. As for the three most common mistakes: first, criteria too vague (for instance just "is it good?"), leaving the AI nothing to measure; second, not requiring reasons, so you get scores without insight; third, not specifying "from a strict / investor perspective," so the AI slides back into pleasing mode. Avoid these three and you will reliably get useful criticism. Once you master the art of asking, you will see that the AI's real value is becoming a dependable member of your team. Visit ai.ud.hk to explore UD's AI Staff solutions and see how AI can share your team's workload.
懂AI，更懂你｜UD相伴，AI不冷
