Spot what's wrong with this prompt: 'You are a helpful assistant. Please be thorough and accurate. Give a good answer.'

This prompt has three directives, and every one of them is a **vague adjective** with no observable target. Helpful, thorough, accurate, and good all sound like quality words, but none of them tells the model anything it can act on. The model cannot check whether its output is thorough enough or good enough because those words do not point at anything measurable. The fix is to walk each directive and replace it with either a concrete action the model can perform or a measurable property of the output. Instead of you are a helpful assistant, name a specific role and scope. Instead of be thorough and accurate, list the testable behaviors you want, like cite sources from the provided context. Instead of give a good answer, give a length cap and a structure. The rewrite forces you to say what good output actually looks like. That is the more useful constraint anyway, and it gives your eval suite something to check. The original three sentences are token padding without any behavioral signal.

Every directive in the prompt fails the operational test. You are a helpful assistant gives the model a vague register cue but no concrete role or scope. Be thorough and accurate names two unfalsifiable properties the model has no way to self-check. Give a good answer is the same failure repeated with a different adjective. The fix is mechanical. Replace each vague directive with one of two operational shapes: an action the model can perform, or a measurable property of the output. You are a Python tutor explaining concepts to beginners gives a concrete role and scope. Cite specific sources from the provided context and flag any unsupported claim is a testable action. Answer in under 200 words; structure as definition, example, common pitfall is a measurable property plus a structural constraint. The three rewrites do two things at once. They give the model targets it can actually hit, and they give the eval suite metrics it can check. Citation rate, format compliance, word count, structural conformance, refusal rate. Each maps to a Promptfoo or Braintrust assertion in two lines of code. The broader pattern is that prompt engineering and evaluation are two sides of the same contract. If the directive is unfalsifiable, the model cannot follow it and you cannot improve it. Refactoring vague to operational is the most consequential edit you can make on a production prompt.