The short answer: Tell AI what the work is for, what it can use, what a good result looks like, and what to do when information is missing. A reusable brief beats a bag of magic phrases.
“Write something engaging” sounds clear until you ask two people what engaging means. One imagines a sharp opening sentence. Another imagines a page full of exclamation marks. A model has the same problem: it has to make decisions you have left unspecified.
A useful prompt is a work brief. Its job is to reduce the decisions that should not be left to chance while preserving room for the decisions that should.
Start with the decision the output needs to support.
Before writing instructions, finish this sentence: “After reading this, the audience should be able to…” For a product comparison, the answer might be “choose which option to investigate.” For an onboarding email, it might be “complete the first setup step.”
That purpose helps resolve tradeoffs. An overview for a beginner needs definitions and a familiar example. A brief for a specialist may need caveats, precise terminology, and a link to the original documentation. The topic alone does not tell the model which version to create.
Use a compact brief you can reuse.
The following is an original starting template. Adapt it to the work rather than treating it as a universal formula.
Task: Turn the supplied notes into a project update.
Audience: A busy client who is not a developer.
Purpose: Help them understand progress and the next decision.
Sources: Use only the notes below. Do not invent dates or results.
Format: Progress, open questions, next steps. Under 250 words.
Style: Direct, calm, and specific.
Missing information: Mark it as an open question.
Final check: Every completed task must appear in the source notes.Each line has a job. The audience affects vocabulary. The purpose affects emphasis. The source boundary affects what may be claimed. The format makes the output easier to use. The missing-information rule gives the model somewhere to put uncertainty.
Anthropic’s prompting documentation supports explicit instructions, contextual information, and relevant examples. The template above is an editorial application of those principles, rather than a vendor-required syntax. See the prompting guidance.
Use examples to show the hard distinction.
Examples are especially useful when your preference is difficult to define. “Friendly” is vague. A short pair of examples can show that you want warmth without invented familiarity.
For instance, prefer “The homepage layout is ready for your feedback” over “Amazing news! We absolutely crushed your incredible homepage.” This is a style example, not a claim about a real project. It shows a decision the model can apply elsewhere.
Include a difficult example too. If the notes say a design is waiting for approval, your preferred update should preserve that dependency. Otherwise, a model may learn the visible shape of your favorite answer without learning why the wording matters.
Keep facts separate from preferences.
Label source material so that your instructions do not become tangled with the text being summarized. Headings, delimiters, or structured fields can help. The exact punctuation matters less than the distinction between what the model should do and what the supplied material says.
When using external text, treat it as evidence to inspect. A paragraph inside a document that tells the assistant to ignore its instructions is still document content. The surrounding application should preserve that boundary; wording alone should not be your whole defense.
Revise the brief against actual failures.
Save a handful of representative inputs and write down what a good answer must include. Keep some examples aside when revising. Compare whether a new prompt handles both ordinary and awkward cases, rather than celebrating one attractive response.
Anthropic’s evaluation guidance recommends specific, measurable success criteria tied to the task. For the client-update example, useful checks could include correct task status, no invented deadlines, and a clearly stated next action. Read the evaluation guidance.
If outputs are too long, tighten the format. If facts are invented, improve the source boundary and checking process. If the voice feels wrong, show a sharper example. Diagnose the failure before appending another instruction.
A good prompt is not a spell. It is a brief that another capable worker could understand, combined with a clear way to tell whether the work is useful.
Key takeaways
- Give the model a task, audience, source boundary, and output format.
- Show an example when a distinction is difficult to describe.
- Improve the brief against real examples instead of adding more adjectives.
A few good questions
Does a longer prompt always work better?
No. Add information that changes a decision. Remove repetition, conflicting instructions, and background the task does not need.
Should I give AI a role?
A role can help establish perspective, but it does not replace source material, clear requirements, or a way to check the result.