Learn With Nathan
Sign in
Full sample chapter · Prompting & Communicating with AI
How to Write a Great AI Prompt
Turn a vague request into a clear work order that gives AI the goal, context, constraints, evidence, success criteria, and output shape it needs.
Chapter 01
How to Write a Great AI Prompt
Turn a vague request into a clear work order that gives AI the goal, context, constraints, evidence, success criteria, and output shape it needs.
“Write me a marketing email” is not wrong. It is simply under-specified. The system must guess the audience, offer, purpose, tone, evidence, length, and next action. Sometimes those guesses look impressive. Sometimes they create a polished answer for the wrong job.
A great prompt is not a magic phrase. It is a compact work order. It explains the outcome, supplies the information that should guide the work, marks the boundaries, and makes the result easy to judge.
The best prompt is also proportional. A two-sentence cleanup needs little instruction. A consequential business analysis needs sources, assumptions, evaluation criteria, and approval boundaries. Prompting is the skill of giving enough structure without burying the task in noise.
Start with the job, not a question
Tell the system what work to complete and why it matters
Begin with an action and an outcome: draft, compare, extract, explain, classify, plan, critique, or transform. Name the audience or user and the decision the output should support.
“Can you help with my proposal?” invites a conversation. “Review this proposal for unsupported claims and missing decision information” assigns a job. The second prompt establishes what the model should do with the material.
Supply the context that changes the answer
Include facts, audience, definitions, and prior decisions
is information the model needs to interpret the job correctly: who the audience is, what has already happened, which terms have special meaning, and which facts are authoritative.
Use labels such as Background, Approved facts, and Source material. Separate evidence from instructions so the model can distinguish what it should follow from what it should analyze.
Do not paste everything available. Relevance matters more than volume. Include a detail when changing or removing it could reasonably change the output.
Define constraints and approval boundaries
Say what must, may, and must not happen
control length, tone, format, scope, sources, prohibited content, timing, and uncertainty. Use precise language: “Use only the supplied policy,” “Do not invent prices,” or “Keep the draft under 180 words.”
For actions, state the . The model may draft a customer message but must not send it. It may recommend a refund but must not issue one. Current official OpenAI guidance similarly emphasizes defining goals, relevant context, constraints, required evidence, success criteria, output format, and action boundaries.
Use a six-part prompt blueprint
A repeatable structure makes missing information visible
The practical prompt blueprint
Goal, context, evidence, constraints, success, and format
Audiobook description
Five boxes across the top define the goal, context, evidence, limits, and success criteria. A sixth format box below specifies the shape and length of the answer. Arrows lead to a final testable-candidate box, where the response is reviewed against the stated criteria.
Test the prompt on a real case and revise the cause
A prompt is good when representative outputs pass the rubric
Run the prompt on several realistic cases. When an answer fails, identify whether the goal, context, evidence, constraint, success criterion, or format was missing. Change one part and rerun the same cases.
A longer prompt is not automatically better. Official OpenAI guidance recommends lean prompts, stating each instruction once, keeping examples when they encode a real requirement, and validating changes on representative tasks.
Keep the improved prompt only if it performs better across more than one example. Prompt libraries should store tested work orders, not long collections of unverified phrases.
Pause and apply
Reflection questions
- Which part of your most common prompt currently forces the model to guess?
- What evidence should control the answer instead of general model knowledge?
- Which three criteria would make the output immediately usable?
Sources and further reading
Chapter 1 endnotes
- OpenAI. Model Guidance — Prompting Best Practices, 2026. Official OpenAI documentation.
- OpenAI. Model Guidance — Favor Leaner Prompts, 2026. Official OpenAI documentation.
- Agent Worker Academy. Prompting for Business Results Production Brief, 2026. Academy blueprint.