Most disappointing AI output starts with an underspecified request. The model is asked to guess the audience, goal, constraints, source material, and definition of success. It fills those gaps with plausible defaults, then delivers something polished that may still be wrong for the job.
A strong prompt does not need elaborate wording or a fictional expert persona. It needs a clear task, enough context, explicit boundaries, and an output you can inspect. Treat prompting as writing a compact brief for a fast collaborator who knows a lot but cannot read your mind.
The best prompt is not the longest one. It is the shortest prompt that removes the important ambiguity.
The five-part prompt brief
For work that matters, include these five elements. You can omit any element the model can safely infer, but make that a deliberate choice.
- Outcome — state the decision, artifact, or change you need.
- Context — explain the audience, situation, and relevant background.
- Inputs — provide the source material and label it clearly.
- Constraints — set scope, tone, length, exclusions, and non-negotiable facts.
- Acceptance criteria — describe the format and what a good answer must demonstrate.
OUTCOME
Create [specific deliverable] so that [audience] can [use it].
CONTEXT
[Relevant background, audience, current state, and goal]
INPUT
---
[Paste or attach source material]
---
CONSTRAINTS
- Use only the supplied evidence for factual claims.
- Keep the response under [limit].
- Do not [important exclusion].
OUTPUT
Return [sections, table, JSON schema, or other format].
QUALITY CHECK
Before answering, verify that every requirement is covered.
If essential information is missing, ask up to three questions.From a vague request to a reliable one
Suppose you ask, ‘Summarize this report.’ The model does not know whether you want an executive briefing, study notes, or a list of risks. A better prompt specifies the reader and the decision the summary should support.
Summarize the attached customer-research report for a product
manager deciding what to build next quarter.
Return:
1. Five findings, each tied to evidence in the report.
2. The three highest-impact user problems.
3. Recommended actions labeled high, medium, or low confidence.
4. Open questions the research does not answer.
Use plain language and stay under 600 words. Do not invent counts,
quotes, or conclusions. If evidence is weak or conflicting, say so.This version is better because its output can be reviewed. A recommendation without evidence violates the brief; an unsupported number is easy to spot; and uncertainty has an explicit place instead of being hidden behind confident prose.
Reusable prompt: turn an idea into a product brief
Role prompts are useful when they activate a relevant perspective, but stacking impressive job titles is not a substitute for requirements. The following prompt uses perspectives only to organize the work.
Turn the product idea below into a decision-ready product brief.
IDEA
[Describe the user, problem, and proposed solution]
Analyze it through four perspectives, in this order:
1. Product: user, painful problem, alternatives, value proposition,
assumptions, smallest testable MVP, and success metrics.
2. Engineering: system boundaries, simplest viable architecture,
major data entities, integrations, security risks, and trade-offs.
3. Experience: primary journey, critical screens, accessibility needs,
empty/error states, and the moment users first receive value.
4. Launch: ideal early adopters, positioning, acquisition experiments,
launch checklist, and feedback loop.
For every major recommendation, include the reason and one risk.
Separate known facts from assumptions. Do not fabricate market data.
End with: decisions needed, top five risks, and a two-week validation plan.Reusable prompt: understand a research paper
Explain the attached paper to a technically curious reader who is
new to this topic. Base every claim on the paper.
Structure the response as:
1. One-paragraph summary.
2. Problem and why it matters.
3. Method, explained step by step.
4. Dataset and experimental setup.
5. Main results, including exact metrics and baselines.
6. What the results do—and do not—support.
7. Limitations, possible failure modes, and unanswered questions.
8. A glossary of essential terms.
Use an analogy only when it preserves the technical meaning. Cite the
paper's section, figure, or table for each important result. If the
paper does not provide a detail, write “not reported.”Use iteration instead of one giant prompt
Complex work is usually more reliable as a short conversation. First align on the plan, then create a draft, then critique it against explicit criteria. This keeps mistakes visible and lets you correct direction before the model produces pages of unusable output.
- Plan — ask for assumptions, missing information, and a proposed approach.
- Draft — provide the missing context and request one concrete deliverable.
- Review — ask the model to check the draft against your acceptance criteria, not to judge itself vaguely.
- Verify — confirm names, numbers, quotations, calculations, links, and high-impact claims independently.
Common mistakes and the smaller fix
- “Act as the world's best expert.” Replace prestige with the specific perspective or standard you need.
- “Be accurate.” Supply trusted sources and require the model to mark unsupported claims.
- “Think step by step.” Ask for assumptions, calculations, evidence, and a concise rationale you can inspect.
- Dozens of conflicting rules. Prioritize requirements and remove instructions that do not affect the result.
- Pasting sensitive information. Redact secrets and personal data, and follow your organization's approved AI policy.
- Trusting fluent output. Treat fluency as presentation quality, not proof of correctness.
A 30-second check before you send
- Could two reasonable people interpret the requested outcome differently?
- Did I identify the audience and the decision this output supports?
- Are source material and instructions clearly separated?
- Did I state what the model must not assume or invent?
- Can I quickly tell whether the response meets the requested format and quality bar?
Prompt engineering is simply careful communication under uncertainty. Give the model a clear destination, the evidence it may use, and a finish line you can verify. Then keep human judgment where it belongs: checking facts, weighing trade-offs, and owning the final decision.