Say what success looks like

Why it matters

When I read a prompt, the first thing I look for is the finish line. A model will always produce something; the question is whether you can tell, afterwards, that it is the thing you needed. If the prompt only names a topic, the model has to guess the outcome, and you have to judge the result by feel.

The evidence points the same way. In the ROPE study, training people to state their requirements completely raised participants' scores from 21.9% to 40.9%, while conventional prompt-engineering training gained 0.7%, and an automatic prompt optimiser could not close the gap. A separate study of journalists found that prompt training made them feel more expert, while its effect on the quality of their work was mixed. Knowing tips is not enough; saying what you need is the skill.

Vendors say much the same. Anthropic's guidance asks for success criteria that are specific and measurable, and OpenAI's advice for reasoning models is to define clearly what a successful response looks like. A goal with a test attached is something both you and the model can check.

What good looks like

• 0 — No discernible outcome. A subject with nothing to do: “Tell me about marketing.”
• 1 — Topic or activity, outcome implicit. You know what you want; the prompt does not say it: “Write about our new product.”
• 2 — Outcome stated, no success criteria. “Write a launch email for our product.” Better, but any email would pass.
• 3 — Outcome plus at least one testable criterion. “…that gets existing customers to book a demo; under 150 words.” Now you can check the result.
• 4 — Criteria with a priority. “…a single call to action above the fold; brevity matters more than covering every feature.” The model knows what to give up when the criteria clash.

Common mistakes

• Naming an activity (“look at”, “help with”, “write about”) instead of an outcome.
• Criteria nobody could test, such as “make it engaging” or “high quality”. Ask yourself how you would recognise engaging.
• Hiding the real purpose: the email is meant to win demo bookings, but the prompt never says so.
• Listing several criteria and leaving the model to guess which matters most when they conflict.
• Confusing length with success. A word limit is a fine criterion, but rarely the only one that matters.

Try it on your own prompt

Take a prompt you used this week and paste it into the workspace. Before you look at the score, finish this sentence for yourself: “I will know the answer is good if…”. If that sentence is not in your prompt, add it, then compare the goal score before and after. The coach may ask you a question or two to help you find the criterion you left out.