AI Prompt Engineering: Examples, Templates, and a Simple Framework

Prompt engineering is the practice of designing and testing instructions so an AI system produces a useful result more consistently. For most everyday work, it does not require special syntax. It requires a clear success condition and a disciplined way to improve the instruction.

The brief, evidence, output framework

Use three layers:

• Brief: What job should the model do, for whom, and why?
• Evidence: What source material, facts, examples, or rules should guide it?
• Output: What exact structure should the result follow?

Then add a check: ask the model to list unresolved assumptions or validate the answer against the brief.

Example: rewrite a landing page section

Weak prompt:

Improve this copy and make it compelling.

Engineered prompt:

Rewrite the hero section for operations managers at 50–200-person logistics companies. Preserve the claim that setup takes one day. Do not add statistics. Use a concrete headline, a two-sentence explanation, and one call to action. Keep the tone confident and plain. After the draft, list any claim that still needs evidence.

The improved version defines the audience, protected fact, prohibited behavior, output shape, and quality check.

Example: research without hiding uncertainty

Research question: [question]
Scope: [region, period, industry]
Use: [decision this research will support]
Required evidence: Prefer primary sources and show publication dates.
Output: Findings, source table, disagreements between sources, evidence gaps.
Rule: Do not convert an estimate into a fact. Label inferences explicitly.

For a deeper starting point, adapt the rigorous research partner.

Example: code review

A useful code-review prompt should identify the language and runtime, describe the intended behavior, provide the relevant diff or files, and rank findings by impact. It should also say what not to spend time on.

Review this TypeScript change for correctness, security, and regressions.
Context: The function runs in an API route and receives untrusted input.
Report only actionable findings. For each finding include severity, file/line, failure scenario, and a minimal fix.
Do not comment on formatting unless it causes a defect.
If you find no issues, say so and list the tests you would run.

The code review specialist offers another reusable structure.

Test prompts like small experiments

Create three to five representative inputs, including an easy case, a typical case, and an edge case. Run the same prompt on each. Score the outputs using a short rubric such as factual accuracy, completeness, format compliance, and usefulness.

Change one major instruction at a time. If you change the role, examples, constraints, and format together, you will not know which change mattered. Save the prompt version beside the test cases so you can compare results after a model update.

When examples help

Examples are valuable when the desired style or classification boundary is hard to describe. Include a small number of representative examples and explain what makes them correct. Do not overwhelm the task with examples that conflict or contain irrelevant details.

Keep the system honest

No prompt guarantees truth. Models can misunderstand context, follow bad source material, or produce confident errors. Prompt engineering improves the odds of a usable result; verification makes the result dependable. The strongest workflow combines a clear brief, grounded evidence, a precise output contract, realistic test cases, and human review.