The fastest way to improve an AI answer is rarely a clever phrase. It is giving the model a finish line it can actually see.
A useful prompt separates three things:
- The job — what decision, draft, plan, or transformation you need.
- The operating constraints — audience, source boundaries, format, length, tone, and things it must not do.
- The acceptance test — the small checklist that tells you whether the result is usable.
OpenAI’s current guidance makes the same split: keep durable role and tone guidance in a system message, put task-specific details and examples in the user message, and treat production prompts like versioned code with tests.[1] Anthropic similarly recommends defining success criteria, a way to test them, and an initial draft before tuning the prompt.[2]
The trick: write a “definition of done”
Instead of this:
Summarize these customer notes.
Use this:
Task: Summarize the customer notes for the product lead.
Audience: A busy PM who has not read the notes.
Output: 5 bullets maximum: recurring problem, affected users, evidence, proposed next step, and open question.
Rules: Do not invent quotes or counts. Flag disagreements.
Done when: The PM can decide whether to investigate this week without opening the source notes.
The second version does not merely ask for a summary. It names the reader, constrains the shape, establishes a truth boundary, and supplies a practical test. That reduces the “almost right” answers that trigger another prompt.
Add a source boundary
For factual work, say exactly what the model may use:
Use only the attached notes. If evidence is missing, write “not established in the supplied material.”
That sentence is more reliable than a vague request to “be accurate.” It gives the model a safe behavior when the evidence is thin: surface the gap instead of filling it.
Ask for a check before the final
For work that will be sent, published, or turned into code, request a short self-check:
Before answering, check the draft against the acceptance test. List any requirement you could not satisfy.
This is not magic. It just makes the desired quality bar explicit. When the task has measurable stakes, save a few representative examples and evaluate prompt changes against them—an approach both OpenAI and Anthropic recommend in their current guidance.[1][2]
Copy this prompt skeleton
ROLE / CONTEXT
You are helping [audience] make [decision or outcome].
TASK
Create [specific deliverable] from [allowed inputs].
CONSTRAINTS
- Format: [structure, length, fields]
- Tone: [tone]
- Evidence: use only [sources]; flag missing evidence
- Do not: [known failure mode]
DEFINITION OF DONE
The output lets [audience] [concrete next action].
It includes: [2–4 observable requirements].
FINAL CHECK
Before sending, compare the result to the definition of done.
State any unmet requirement plainly.
Use it today
Take the prompt you used most recently and add only two lines:
- “Done when…”
- “If you cannot verify a claim from the supplied material, say so.”
Those two lines often do more than adding another paragraph of background. The point is not to make prompts longer. It is to make the target unambiguous.