Better Prompts Begin with Clear Success Criteria

Environment and inputs: An approved AI assistant, one recurring task, three representative inputs and a written answer-checking rubric.

Repeatedly rewriting a prompt is difficult when nobody has defined success. Anthropic recommends establishing measurable criteria and a way to test them before tuning prompts. A better instruction describes the intended audience, available evidence, required format, and what to do when information is missing.

What matters

• State the exact task and provide only the context needed to complete it.
• Show a representative example of the output you want, including important constraints.
• Compare revisions against the same small set of realistic cases instead of relying on one good answer.

Practical example: For a handover summary, request the issue, actions taken, remaining risks, and next owner in four labelled sections.

One action today: Choose one recurring prompt and define three criteria for checking its answers.