Prompt Quality

What it measures

A prompt without a testable Definition of Done is a guess. Most correction rounds are symptoms of underspecified prompts, not model errors.

Before sending a prompt, can you state exactly what the output must satisfy for you to accept it?

The four diagnostic questions

1. 01Each prompt contains only the context relevant to that specific task — not the full data model or all rules.
2. 02Prompts express business logic as explicit conditions (if X then Y, must not Z), not vague prose.
3. 03Every prompt has a testable Definition of Done. You know before sending whether the output will be acceptable.
4. 04You batch related changes into single prompts with ordered steps rather than sending incremental one-liners.

What it looks like when it works

Strong prompt discipline. Consider whether batching could further reduce round-trip overhead.

What goes wrong

Prompts are too broad. Scope context to the task, express logic as conditions not prose, and define what done looks like before sending.

Critique & references

Calls out negation blindness — the structural asymmetry where prohibitions are processed less reliably than affirmative instructions. Explicit conditional logic (if/then/must-not) reduces parsing ambiguity more than tone or politeness ever does.

• → Anthropic — Be clear, direct, and detailed
• → Negation in language models — survey (arXiv 2309.02747)