Claude Coding

6 Prompts to Evaluate an AI Model Before You Commit

Vendor benchmark cards do not tell you whether a model works on your codebase. Two models released this month published scores on different versions of the same benchmark, which makes their headline numbers meaningless against each other. These six prompts are what we run instead: each one is designed to fail in a way you can see, rather than to produce output that looks impressive. Run the same six against every model you are considering and compare the failures, not the successes.

⌨️ 6 prompts 🕐 Updated Oct 12, 2026

💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.

The Admission Test

Reveals whether a model can separate what it knows from what it is guessing. A model that writes nothing under CONFIDENCE on a hard question is not calibrated.

Answer the following question. Then, in a separate section headed CONFIDENCE, state which parts of your answer you are least sure about and what would change your mind. Do not hedge inside the answer itself. Question: [PASTE YOUR QUESTION]

The Stale Fact Trap

Pricing changes constantly. A model that gives a confident number with no date caveat will do the same thing on every other fact it half remembers.

What is the current pricing for [TOOL NAME]. If you are not certain this is current, say so explicitly before giving any figure, and state the date your information is from.

The Real Codebase Test

Asking for a rewrite rewards fluent output. Asking for failure modes rewards understanding. The difference between models is much larger on the second.

Here is a function from our codebase. Do not rewrite it. Instead list, in order of severity, every way it could fail in production, and for each one state the input or condition that triggers it. [PASTE FUNCTION]

The Contradiction Check

Weak models split the difference or pick one silently. Strong models locate the disagreement precisely and say what it would take to resolve it.

Read the two sources below. They disagree. Identify exactly where they conflict, state what evidence would settle it, and tell me which one you would act on and why. Do not average them. [SOURCE A] [SOURCE B]

The Refusal Boundary

Tests whether a model will argue against its own use. Anything that claims it can do every part of every task is telling you nothing useful about scope.

I need you to do [TASK]. Before you start, tell me which parts of this you cannot do well, and whether any part of it is better done by a tool other than you.

The Long Context Probe

Advertised context windows and usable context windows are different numbers. Forcing a quote from the far end of the document is how you find out which you are getting.

I am pasting a long document. Without summarising it, answer this single question using only information from the second half of the document, and quote the exact sentence you relied on. [PASTE DOCUMENT] Question: [YOUR QUESTION]