What it is
A description of a good result is a short set of success criteria written into the prompt. On the ten checks this is Standard, the seventh check. The criteria worth writing are ones someone could check by reading the answer: what must appear and where, what the reader must be able to do afterward, what must be true of every item in a list.
Adjectives are not criteria. “Professional”, “engaging” and “high quality” each mean something different to every reader, and the model has to guess which meaning you had. Ask yourself how you would tell a good answer from a weak one if three landed on your desk, and write that down.
When it helps
It helps when “good” is not obvious from the task. A list of action items is good if it is complete; a persuasive email is good if the reader agrees to the meeting; a summary for a decision is good if the decision can be made from it. Those are different targets, and a model that knows which one it is aiming at makes different choices about what to include.
It helps when you are asking for a rewrite, because “make it better” has no direction. Say what better means here: shorter, answerable on a phone, free of jargon a customer would not know.
And it helps most with prompts you will reuse. Criteria you write now become the checks you run later. The guide to testing prompts turns each one into a pass or fail rule, and Prompt Tester runs them on real inputs.
When it does not
Criteria nobody could check add words without adding direction. If you cannot say how you would tell whether the answer met a criterion, cut it or make it concrete.
Too many criteria pull against each other. “Short” and “complete” cannot both win when the material is long. Pick the two or three that matter most, or rank them, so the model knows what to give up first.
The model may tell you it met the criteria when it did not. Read the answer against them yourself, or ask in a second turn for each criterion to be checked with a quote from the answer, the approach in draft, then critique.
Criteria are not facts either. “States the payback period” only works if the payback period is in the material you pasted, so pair this with giving it the facts. For a quick one-off question, skip all of it: you will judge the answer when you see it.
Before and after
Three requests where the after version describes what good looks like. Each score comes from the Prompt Grader’s free ten-check estimate, written for any chat model.
Write an executive summary of the pilot proposal below for our finance director. """ [the proposal, pasted in full] """
Write an executive summary of the pilot proposal below for our finance director. A good summary lets them decide in two minutes whether to fund the pilot. It states the cost, the payback period and the one risk that could change the answer, all in the first paragraph. """ [the proposal, pasted in full] """
Without a target, an executive summary can shrink the proposal section by section, background first. With one, the model leads with the three numbers a funding decision turns on and can drop the history.
Give me feedback on the slides below before I present them to the leadership team. """ [the slide text, pasted in full] """
Give me feedback on the slides below before I present them to the leadership team. A good review names the three changes that would most improve the deck, each tied to a slide number, and says what to keep as it is. Comments on wording only count if the wording would confuse someone. """ [the slide text, pasted in full] """
Open-ended feedback can come back as a long list of small edits with the one real problem somewhere in the middle. The second version gets three ranked changes you can make before the meeting, plus what not to touch.
Rewrite our help-center answer about resetting a password, below. """ [the current help-center answer] """
Rewrite our help-center answer about resetting a password, below. A good answer gives the first step in its first sentence, works for customers on both the web app and the phone app, and leaves no reason to contact support. """ [the current help-center answer] """
“Rewrite” on its own can mean “reword”. The criteria turn it into a fix: the first step moves to the top, and a model that has to cover both apps will notice if the current answer only describes one.
Each after version gains three points. Standard is one; the criteria also set a size or name the reader, and the separate paragraph passes Fit.
The line to add
A good result lets {{reader}} {{do_what}}. To pass, it must {{first_check}} and {{second_check}}.
Write the first sentence as an outcome for a person and the second as things you could tick off. If you only write one, write the outcome: “lets a new hire set up their laptop without asking anyone” tells the model more than a page of adjectives.
In ChatGPT, Claude, Gemini and Copilot
OpenAI is the most direct about this. Its guide to its latest models, the ones behind ChatGPT, says that even though the model can infer much of the work from context, you should still give it success criteria. The same guide says broad tone labels such as “friendly” or “empathetic” can be ambiguous, and recommends describing the writing choices that make up the tone instead, which is the same move as replacing adjectives with checks.
Google’s guide for Gemini 3 asks you to explain any ambiguous terms or parameters explicitly. Anthropic’s page on defining success criteria is written for testing, and its advice carries over to the prompt: make criteria specific and measurable, so “classify sentiments well” becomes a target you can score. In Microsoft’s four parts of a Copilot prompt, criteria belong with expectations, how Copilot should respond.
Key takeaways
- Write criteria someone could check by reading the answer.
- Swap adjectives such as “engaging” for what the reader can do after reading.
- Keep to two or three criteria, ranked if they could conflict.
- Check the answer against the criteria yourself; the model’s own claim is not proof.
- For a prompt you reuse, the criteria become your test checks.
Questions
What are success criteria in a prompt?
A short description of what a good answer must do, written so someone could check it: “states the cost and the deadline in the first paragraph” or “every action item has an owner.” They tell the model what to aim for and give you a way to judge the result. Vague words such as “high quality” are not criteria, because nobody could test them.
How do I get better quality answers from ChatGPT?
Say what better means for this task. Name the reader and what they must be able to do after reading, give the facts the answer depends on, and add two or three criteria you could check. Then read the answer against those criteria and ask for a revision on the one it missed. The ten checks list the rest.
Should I ask AI to check its own work?
It helps, as long as the check is specific. Ask it to go through each criterion and quote the part of the answer that meets it, or say that it does not. A general “check your work” can get a general “looks good.” Read the result yourself before you rely on it.
Prompts to use with this
- Give specific feedback to someone on your teamYou need praise or correction that names the behavior, its impact and what to do next.
- Screen applications against a role's must-have criteriaYou have a pile of applications and want each checked against the same job-related criteria, with a person making every decision.
- Check a draft for language that could exclude readersBefore a document goes to a wide audience, you want wording that could exclude, stereotype or lose some readers flagged, with plain alternatives you can accept or reject.