Describe what a good result looks like

“Make it clear and professional.” “Make it engaging.” “Make it good.” Almost every prompt wants this, and almost none of it changes the answer, because the model was already trying to be clear and professional. Those words are wishes, not instructions.

A standard is different. It is a test you could apply to the result: “After one read, the customer knows what happened, what we are doing about it and when they will hear from us.” Hand that sentence to a colleague and they could tell you whether a draft passes.

Why a test works better than an adjective

The model makes many small choices as it writes: what to put first, what to leave out, how much to explain. An adjective gives it no way to choose between them. A test does. If the standard is “a new agent can follow it on their first day without asking anyone,” the model knows to spell out each click and skip the history of the policy.

Write a good, clear summary of last night's outage.

Write a summary of last night's outage. A good summary lets someone who was not on call understand in one minute what broke, which customers were affected and what we changed so it does not happen again.

“Good, clear” gave the model nothing to aim at; the new sentence is a test a reader could apply, and “in one minute” also sets how short it must be.

Tone words have the same problem. Friendly to one reader is too casual to another. Describe the choices that make the tone instead: “short sentences, we and you, and one line that admits the change is a hassle.” OpenAI’s guidance suggests the same swap, from broad tone labels to the concrete writing choices behind them.

Ways to write a standard

A standard can name what the reader can do afterward, what they should come away knowing, a check you could run, or who would sign the result off:

• “A new hire can set up their laptop from this alone.”
• “The client understands the delay was ours and knows the new date.”
• “Every figure matches the spreadsheet, and every action has an owner.”
• “Our legal team would approve it without edits.”

A standard keeps working after the answer comes back. You can check the draft against it yourself, or ask the model to check its own draft against it before you read it. When a prompt will be reused, the standard becomes the start of a test; the guide to testing prompts shows how. The describe a good result pattern has more phrasings.

One honest note about the score box: the free estimate looks for signs of a standard, so a phrase like “make it clear” can pass it. Aim for a test a person could apply, not for the word the estimate notices.

If you cannot say how you would judge the result, the model cannot aim for it. Write the standard first, then the rest of the prompt.

Write a step-by-step guide to issuing a refund in our billing system, using the notes below. New agents will use it during their first week. No more than 8 steps. Leave out partial refunds, which have their own guide. A good guide is one a new agent could follow on their own the first time, without asking anyone for help.

Notes:
"""
Open the customer in Billing. Find the invoice. Click Refund (top right). Choose Full refund. Pick a reason from the list. Add the ticket number in the note field. Click Confirm. The customer gets an automatic email. Refunds over $500 need team lead approval first.
"""

Summarize the attached travel expense policy for our managers, who will use it to check their team's claims before approving them. Fit it on one page, as a table with one row per expense type and columns for the spending limit and the receipt needed. Leave out the booking rules, which managers don't handle. A good summary lets a manager decide in under a minute whether a claim is within policy, without opening the full document or emailing finance.

Write an email to a candidate who reached the final round for our data analyst role but did not get the job. We chose someone with more SQL experience. The panel liked their presentation and would consider them for a junior analyst opening we expect in the fall. Under 150 words, with a subject line. Don't mention the other candidate. A good email leaves them knowing the decision is final, why it went the other way, and that the invitation to apply again is real.

Remember. Replace adjectives with a test: what the reader can do, know or decide after reading.