Which rewrite of the directive 'Be thorough and accurate in your response' is most likely to actually change model behavior?

**Vague adjectives** in prompts feel strong but do not give the model anything to act on. Words like thorough, accurate, helpful, and good have no observable target. The model cannot check whether its output is thorough enough or accurate enough, because those words do not point at any property it can measure. The fix is to replace each vague adjective with a concrete action or a measurable property. Instead of be accurate, write cite at least one source from the provided context for every factual claim. Instead of be thorough, write structure your answer as definition, example, common pitfall. Instead of be good, write keep your response under 200 words. Each rewrite gives the model a target it can hit and gives you something the eval suite can check. The original adjectives gave you neither. Stacking emphasis on vague adjectives (PLEASE be VERY thorough) does not fix the mechanism, it just spends more tokens for the same un-actionable instruction.

Every directive in a prompt should map to either an observable action the model can perform or a measurable property of the output. Vague adjectives like thorough, accurate, helpful, and good fail both tests. The model has no way to check itself against them, and your eval suite has no way to score them. The rewrite is mechanical. Replace each adjective with a concrete rule the model can follow. Cite at least one source from the provided context for every factual claim is testable. Keep your response under 200 words is measurable. Structure your output as a JSON object matching this schema is checkable. Each turns an unfalsifiable preference into a verifiable behavior. The alternatives all fail the same test in different ways. Stacking emphasis (PLEASE be VERY thorough) does not change the operational content, just the token cost. Repeating the directive does the same. Moving it into the role definition (you are a thorough and accurate assistant) relocates the vague language without operationalizing it. The broader principle is that prompt engineering and eval are two sides of the same contract. If the directive cannot be evaluated, the model cannot be evaluated against it. Concrete directives create checkable behaviors; vague ones create the illusion of control.