Create a practical prompt-writing guide for [WORK TASK] used by [INTENDED USERS]. The AI should produce [REQUIRED OUTPUT], using these available inputs: [AVAILABLE INPUTS]. The main accuracy, confidentiality, or quality risks are [RISKS AND BOUNDARIES]. Write the guide for people who can do the work but are not prompt engineers. Make it 900 to 1,200 words, with these sections: 1. What a good result looks like: three observable criteria tied to this task. 2. What to gather before prompting: a checklist that separates required source material from optional context. 3. A fill-in prompt template: include only necessary [PLACEHOLDERS], specify the output format, and tell the model to ask up to 3 questions only when a required input is absent. 4. An annotated example using a small fictional scenario, with notes explaining why each instruction matters. 5. A revision sequence: three exact follow-up prompts for fixing scope, evidence, and format problems. 6. A final review checklist the human must complete before using the result. Teach specific controls rather than vague prompt advice. For example, explain when to provide source text, when to define an audience, how to set a length or format, and how to require an evidence boundary. If the task involves regulated, legal, medical, financial, personnel, or safety decisions, state clearly what requires qualified review. Do not claim that prompting guarantees accuracy. Before answering, self-check that the template can be completed with the listed inputs and that every review check addresses a stated risk. Ask up to 3 clarifying questions only if a required input is missing.
Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [WORK TASK] | Name the repeatable work task the guide should cover. | Drafting monthly customer-success account reviews |
| [INTENDED USERS] | Describe the people who will use the guide and their working context. | Customer success managers at a B2B software company |
| [REQUIRED OUTPUT] | State the deliverable the AI should produce for this task. | A one-page account review with outcomes, risks, next steps, and a meeting agenda |
| [AVAILABLE INPUTS] | List the documents, facts, examples, and constraints users can provide. | CRM notes, usage metrics, open support tickets, renewal date, and previous review |
| [RISKS AND BOUNDARIES] | List mistakes to prevent, confidential material rules, and any required human approval. | Do not invent product usage or customer commitments; remove personal data; manager reviews churn-risk recommendations |
How to use
- Choose one narrow task instead of trying to document every use of AI at work.
- List the actual source material users can access, including what they must not paste into an AI tool.
- Paste the prompt and test the resulting template on one completed real-world example.
- Follow up with: “Rewrite the template for a user who has only [LIMITED INPUTS], and show what it cannot safely infer.”
Variations
Beginner cheat sheet
Use this when the audience needs a one-page reference rather than a full guide.
Create a one-page prompt-writing cheat sheet for [TASK] for [USER TYPE]. Include a five-part formula, one fill-in prompt with [INPUTS], three common failure modes, and three exact repair messages. The target output is [OUTPUT]. Keep examples grounded in the task and explain what source material must be supplied. Do not promise accuracy or suggest inventing missing facts. Ask up to 3 questions only if a required input is missing.
Team playbook
Use this for a shared team standard with review ownership.
Write a prompt-use playbook for [TEAM] performing [TASK]. Define approved inputs, prohibited inputs, the standard output format [FORMAT], required human reviewer [REVIEWER], and escalation cases [ESCALATION CASES]. Include a reusable prompt, version-control rules for templates, a QA checklist, and five example follow-up instructions. Flag where AI output must be verified against source records. Ask up to 3 questions only when a required input is missing.
Prompt audit
Use this to improve an existing prompt that gives inconsistent results.
Audit this prompt for [TASK]: [CURRENT PROMPT]. The expected output is [EXPECTED OUTPUT], users have [INPUTS], and failures have included [FAILURES]. Diagnose ambiguous instructions, missing context, unsafe assumptions, and output-format gaps. Then provide a revised prompt, a short explanation of each material change, and three test cases with pass criteria. Do not add facts unavailable to users. Ask up to 3 questions only if needed.
Tips
- Build guides around a single observable deliverable, such as a renewal review or incident summary, rather than a job title.
- Treat source material as an input contract: tell users exactly which facts the model may rely on and which it may not infer.
- Include repair prompts for common failures, because first drafts often need narrower scope or a different output structure.
- Test a template using thin inputs as well as rich inputs; this reveals where users need a question, not a confident guess.
FAQ
What is the most important part of a work prompt?
The input boundary is often most important. A model needs to know which documents and facts are authoritative before it can make a reliable output.
Should a prompt ask the AI to role-play an expert?
It can set useful expectations, but it does not supply evidence. Specify the actual task criteria, sources, and review checks instead.
How often should a team update a prompt guide?
Update it after a meaningful process, policy, or source-system change, and whenever repeat users find the same failure mode.