Home / Techniques that change the result

Checking a prompt the way you check a result

An eval is a fixed set of tests that scores a prompt, and meta-prompting uses Claude to improve an existing prompt, but in both cases the result stays a draft to test, never a final answer.

A prompt that has answered correctly three times in a row on the examples you happened to try has not been checked, it has only been tried by chance. An eval changes this situation: it is a fixed set of test cases, written in advance, that automatically scores a prompt on each of them, exactly as a test suite scores a programme.

What an eval actually measures

Claude's documentation describes this cycle in one clear sentence: building a successful application starts with defining success criteria, then designing evaluations that measure performance against those criteria, a central cycle of prompt engineering (platform.claude.com, section on testing and evaluation, accessed 2 September 2026). A minimal eval fits in three lines.

cas 1 : entree "Le colis est arrive casse" -> attendu "reclamation"
cas 2 : entree "Merci, tout est parfait" -> attendu "compliment"
cas 3 : entree "Quel est le prix du modèle B" -> attendu "question"

Run your classification prompt on these three inputs and compare each output obtained to the expected one. A score of two correct answers out of three does not say the prompt is generally good or bad, it says which of the three cases fails, and that is the one to fix first.

Meta-prompting writes a draft, not an answer

Meta-prompting means asking Claude to improve an existing prompt rather than rewriting it alone from a blank page. The Metaprompt tool in the recipe collection published by Anthropic is designed to solve this blank page problem and give a starting point to iterate from, and its documentation says so without hedging: the prompt obtained at the end is not guaranteed optimal in any way, do not hesitate to change it (Anthropic cookbook, Metaprompt page, accessed 2 September 2026). A prompt that comes out of meta-prompting therefore carries no more authority than a first draft written by hand, it must go through the same fixed set of cases before being adopted.

The same set of cases, before and after

The rule linking the two halves of this technique fits in one sentence: do not improve a prompt without retesting it afterwards on exactly the same set of cases that was used to judge it the first time. Changing the test cases between the two passes amounts to comparing two prompts on two different problems, which proves nothing about actual progress. The figure alongside shows this full cycle. This principle connects with test the checker before trusting its verdict: an eval is also a check, and a check that changes cases between two measurements no longer measures the same thing. It carries straight through into asking several times, and contradicting yourself, which applies another form of stress test, this time to a single answer rather than to a whole prompt.