In this guide
  1. Use one prompt structure for every task
  2. Prompt 1: Turn notes into a decision-ready summary
  3. Prompt 2: Draft a clear workplace message
  4. Prompt 3: Prepare for a meeting
  5. Prompt 4: Compare options against explicit criteria
  6. Prompt 5: Challenge a plan before execution
  7. Prompt 6: Improve a draft without changing its facts
01

Use one prompt structure for every task

Build prompts from six parts: task, audience, context, source material, output and limits. State what good looks like and what the system must not invent. Ask it to mark missing information rather than fill gaps.

Use only information approved for the tool and task. Remove confidential, personal or regulated data unless your organisation has explicitly authorised the system and workflow. Treat every result as a draft that remains your responsibility.

Base prompt

Task: [what to do]. Audience: [who will use it]. Context: [what matters]. Use only: [approved material]. Output: [format, length and sections]. Limits: do not invent facts; mark uncertainty [CHECK]. Human review: I will verify [critical items].

02

Prompt 1: Turn notes into a decision-ready summary

Use this when approved notes are accurate but unstructured. Tell the system which distinctions matter: confirmed facts, open questions, risks, decisions and next actions. Require traceability back to the supplied notes.

Review whether an important exception disappeared during compression. A shorter summary is useful only when it preserves the information the reader needs to act.

Using only the approved notes below, create a one-page summary for [audience]. Use headings for confirmed facts, risks, decisions needed and next actions with owners. Do not infer causes or commitments. Place [CHECK] beside missing dates, owners or evidence.

03

Prompt 2: Draft a clear workplace message

Use this for a first draft of an email, update or announcement when you already know the facts and required action. Provide the relationship, tone and channel. Ask for a clear subject or opening and a visible next step.

Check every name, date, promise and implication. Rewrite language that does not sound like you. Never let the model invent a commitment on behalf of another person.

Draft a concise [email/update] for [audience]. Purpose: [purpose]. Include only these verified facts: [facts]. The reader should [action]. Tone: direct, respectful and plain. Keep it under [length]. Do not add promises, reasons or deadlines not supplied.

04

Prompt 3: Prepare for a meeting

Use AI to organise your preparation, not to decide the meeting’s human dynamics. Supply the purpose, participants, permitted background and decision boundary. Ask for questions, assumptions to test and a proposed sequence.

Remove generic agenda items. Verify that the plan gives participants enough context and a genuine route to the required decision.

Help me prepare a [duration] meeting with [participants]. Desired outcome: [decision or alignment]. Context: [approved facts]. Produce: a five-item agenda, three questions that test assumptions, information to send beforehand and a closing recap. Flag any missing input.

05

Prompt 4: Compare options against explicit criteria

Use this when the options and evidence are known. Define the criteria, their importance and any hard constraints. Ask the system to keep ‘not provided’ separate from a negative result.

Inspect the comparison for hidden weighting and unsupported recommendations. The final trade-off and decision belong to the accountable person or group.

Compare options A, B and C using only the evidence below. Criteria: [criteria]. Hard constraints: [constraints]. Return a table showing evidence, gaps and trade-offs, followed by questions that must be answered. Do not rank an option when the evidence is insufficient.

06

Prompt 5: Challenge a plan before execution

Ask the system to examine assumptions, dependencies, failure points and missing stakeholders. This can widen your review, but it is not an independent risk assessment. Give it the actual plan and context instead of asking for generic criticism.

Classify the response into relevant, already controlled and unsupported. Discuss material risks with the people who understand the work and have authority to act.

Review this plan as a constructive challenger. Identify assumptions, dependencies, affected stakeholders, failure signals and questions for the owner. Use only the plan and context supplied. Separate evidence-based concerns from possibilities that need validation.

07

Prompt 6: Improve a draft without changing its facts

Use this when the substance is approved but the structure or language needs work. State which facts, terms and commitments must remain unchanged. Ask for a revised version and a short list of material edits.

Compare the result line by line where accuracy matters. If the system changes meaning while improving style, restore the verified wording. Save successful prompts with their purpose, allowed inputs and review checklist rather than as context-free tricks.

Revise the draft for [audience] using plain language and clear headings. Preserve all figures, dates, names, defined terms and commitments exactly. Do not add claims. Return the revision, then list any sentence whose meaning may have changed so I can verify it.

Before you move on

  • The prompt names the task, audience, context and output.
  • Only permitted information is included.
  • Authoritative sources and hard limits are explicit.
  • The model is instructed to mark gaps and uncertainty.
  • Facts, commitments and consequential recommendations receive human review.
  • The output is checked for omissions as well as inventions.
  • Reusable prompts include their purpose and review requirements.

Sources and further reading

These official resources informed the guardrails and practical method in this guide.

  1. OpenAI Academy: Power Prompting ↗
  2. OpenAI Help: Prompt Engineering Best Practices ↗
  3. NIST Generative AI Profile ↗
  4. CISA: Using AI Securely ↗
Use the free Workweek Starter →

This preview uses fictional examples where indicated. Read our editorial approach.