A weak prompt asks for a document. A strong prompt defines the work the document must do. That difference matters more than clever phrasing, because fluent output can still be operationally useless.
Start with the decision behind the deliverable
Most prompt requests arrive as artifacts: write an email, summarize this meeting, research this company, build a strategy. The form sounds specific, but the purpose is still hidden. An email might need to earn a reply, protect a relationship, document a decision, or force clarity about a next step. Each purpose demands different evidence and a different standard of success.
Before writing the instruction, complete one sentence: ‘This output should help someone decide or do…’ That sentence becomes the operating objective. It tells the model what to prioritize, gives the user a way to review the result, and makes it easier to delete attractive material that does not improve the outcome.
The prompt is not the work. It is the contract for how the work should be done.
Make the evidence visible
Models are rewarded for producing a coherent answer. Operators are responsible for knowing whether the answer deserves belief. A useful prompt must protect that distinction. It should tell the model which statements are supplied facts, which are interpretations, which are hypotheses, and what remains unknown.
This is especially important when the source material is incomplete. Instead of inviting the model to make the brief feel finished, ask it to mark [NEEDS INPUT], attach confidence labels, and identify the missing evidence that could change the recommendation. The resulting document may look less certain. It will be more honest and more useful.
Specify a method, not a personality costume
‘Act like a world-class expert’ is easy to write and difficult to evaluate. A method is more valuable. Ask the account researcher to separate verified signals from assumptions, map each signal to a plausible priority, and write a question that could disprove the hypothesis. Ask the decision facilitator to compare reversibility, downside, values, and information gaps.
The method creates inspectable work. A reviewer can see whether each step happened and challenge the logic where it breaks. The role still matters, but it should establish responsibility and judgment—not theatrical status.
End with a quality check and a next move
A strong output contract names the sections, level of depth, and form of the answer. The stronger version also asks the model to inspect its work before returning it: trace claims to inputs, remove generic advice, surface the strongest counterargument, and check that the result serves the original decision.
Finally, force the work back into the world. Ask for the most important fact to verify and one action that can be taken in the next 24 hours. That closing move keeps the prompt from becoming a content machine. It makes it part of an operating loop: context, judgment, action, evidence, correction.
One useful next step: Choose one idea from this note and test it at the smallest scale that could teach you something this week.