Cold Grader Prompt for Error Detection

A prompting pattern that improves accuracy on hard tasks by using a separate grader that only sees the problem and candidate answer, not the reasoning.

🤖 Works with: Any

The Prompt

Copy and paste — replace anything in [brackets].

You did not write the answer below. You are reviewing it cold. Do not assume it is correct. Identify the single step or claim most likely to be wrong, and say why. Then score from 0 to 7 how likely this answer actually solves the stated problem, where the score reflects whether it solved it, not whether it reads well.

What it’s good for

Improve accuracy on multi-step reasoning tasks like research questions, math problems, or tricky code specs by having a separate model critique answers without seeing the original reasoning.

How to use it

1. Run the initial model to generate an answer to the problem.
2. Use the grader prompt with only the original problem and the candidate answer, withholding all reasoning.
3. Take the critique, generate a fresh attempt with it attached, and repeat for 2-3 rounds, keeping the highest-scoring version.

Curated from the community via Reddit.