AI prompting guide: how to write prompts that stay useful

01

Start with the work, not a magic phrase

Prompting becomes easier when you stop searching for a perfect opening sentence. Begin by naming the work someone must complete. “Help with marketing” is a topic. “Compare three campaign concepts for a new service and recommend the smallest test” is a task with a decision attached.

A clear task gives the model a direction and gives you something concrete to review. Before drafting the prompt, write one sentence describing what the response should help you decide, create, understand, or change.

• Name the user or decision-maker.
• Describe the outcome, not only the subject.
• State what a successful response makes possible.
• Remove requests that belong to a different task.

02

Use six building blocks

Most dependable prompts contain six elements: role, task, context, audience, constraints, and output. The role suggests a perspective, but it does not grant the model real credentials. The task defines the job. Context supplies the facts. Audience shapes language and depth. Constraints prevent unwanted behavior. Output defines the form you can use.

You do not need every block for a simple request. Add detail where ambiguity would change the result. A five-line prompt with strong context can outperform a page of generic instructions.

03

Give the model source material

Specific source material is usually more valuable than more instructions. Provide the notes, page, data, policy, examples, or code that the answer must use. Explain which source is authoritative when several sources disagree.

Use delimiters or labeled sections so the model can distinguish your instructions from the material being analyzed. For long inputs, state whether it should summarize, compare, extract, transform, or critique. Never assume that fluent output means the source was interpreted correctly.

• Label sources and their dates.
• Separate instructions from quoted material.
• Tell the model not to add facts beyond the sources.
• Ask it to identify missing or contradictory evidence.

04

Control uncertainty explicitly

Models often produce a plausible answer when information is missing. Prevent that behavior by deciding what should happen under uncertainty. The prompt can request clarifying questions, labeled assumptions, alternative interpretations, or a list of information still required.

Use different rules for different risks. A brainstorming prompt can tolerate assumptions when they are labeled. A legal, financial, medical, security, or safety-sensitive task requires authoritative sources and qualified human review.

05

Specify a usable output

A format should match the next action. A decision may need a comparison table and recommendation. A meeting may need decisions, actions, owners, and open questions. A content brief needs headings, evidence, and review criteria.

Avoid asking for structure merely because it looks organized. Too many headings can hide weak reasoning. Ask for the smallest form that helps a person use or evaluate the answer.

• Define required sections.
• Set a length only when it serves the use case.
• Request evidence beside important claims.
• Include acceptance criteria or a review checklist.

06

Iterate from evidence

Do not rewrite the entire prompt every time an answer disappoints you. Identify the failure: missing context, vague task, wrong scope, unsupported fact, unsuitable format, or weak example. Change the smallest instruction that addresses that cause.

Keep the prompt and a representative test input together. When the model or workflow changes, rerun the same test and compare results. This is how a prompt becomes a maintained work asset rather than a forgotten piece of text.