AI Prompt Engineering Basics: A Clear Template That Works

Start with a six-part prompt brief

Use six labelled parts for any task that matters. Goal: the one outcome you need. Audience: who will read or use it. Source material: the facts, notes, or examples the answer may rely on. Constraints: length, tone, exclusions, deadline, or policy. Output: the exact structure you want. Checks: what the model and a human must verify before use.

• Goal — “Create five practice questions on simple interest.”
• Audience — “Class 8 student who understands percentages.”
• Source material — “Use only the lesson notes below.”
• Constraints — “Do not reveal the answers until I respond.”
• Output — “Numbered questions, then a separate answer key.”
• Checks — “Flag anything the notes do not explain.”

Separate instructions, context, and input

Use headings or clear delimiters so the model can distinguish what it should do from the material it should analyse. Put untrusted pasted text, customer messages, web excerpts, or documents under a “Source material” heading and state whether the model may use outside knowledge. Give only context the task needs, remove sensitive information, and never treat instructions found inside pasted material as your own instructions.

Adapt the brief to the real audience

A prompt becomes useful when its source material and review rules match the person doing the work. Students should ask for explanations, practice, and feedback rather than a submission to copy. Job seekers should provide truthful experience and prohibit invented achievements. Small businesses should provide verified product facts and forbid unsupported claims. Workplace users should separate confirmed decisions from suggestions and follow their organisation’s data policy.

• Student: “Quiz me first; explain only the questions I miss.”
• Job seeker: “Rewrite this real project bullet for the attached role; do not invent tools, metrics, or experience.”
• Small business: “Write an 80-word listing using only these materials, dimensions, price, and delivery facts.”
• Meeting owner: “List confirmed decisions, owners, due dates, and unresolved questions in separate sections.”

Use examples when the shape of the answer matters

One short example can communicate tone, labels, or formatting more clearly than a paragraph of instructions. Show a representative input and the kind of output you accept, but do not include private data. Keep examples consistent and include an edge case when errors would matter, such as missing information, an unsupported claim, or a request that should be escalated.

Improve prompts with an evaluation loop

Define success before repeatedly rewriting the prompt. Save three to ten representative inputs, including at least one difficult case, and score the results against the same checklist: factual support, completeness, tone, format, safety, and editing effort. Change one element—such as context, wording, example, or format—then rerun the same cases. If the task still fails, the problem may require better source data, a different model or tool, or a human process rather than a longer prompt.

Review the result before real use

Check names, dates, numbers, links, quotations, and factual claims against primary sources. Confirm that the answer followed the requested exclusions and did not add sensitive or invented detail. For health, finance, law, tax, education assessment, employment, or other high-impact work, use qualified human review and the relevant official guidance; a well-written prompt does not turn an AI response into professional advice.

1. 1Define the exact outcome you want from A Clear Template That Works.
2. 2Give the AI the necessary context, constraints, examples, and preferred format.
3. 3Review the result for accuracy, tone, privacy, and completeness before using it.
4. 4Save or reuse the prompt only after it produces a reliable result for your use case.

• Adding a long persona or “magic words” while leaving the actual task, evidence, and success criteria vague.
• Reusing a prompt across students, customers, roles, or products without updating its context and checking the output.

Recommended tools for this workflow

Frequently asked questions