10 Prompt-Writing Techniques for Better AI Results

Better prompts do not depend on clever phrases. They give an AI model the information needed to complete a task: a clear goal, relevant context, usable source material, constraints, and a definition of a good result. The ten techniques below can be combined, but each one solves a different problem.

1. State the task and success criteria

Begin with the action you want and describe what a successful answer must include. “Help with a presentation” is vague; “Create a six-slide outline for managers, with one decision per slide and a final recommendation” is testable.

Task: Compare the two proposals in the attached document.
Audience: Department managers.
A good answer must identify cost, implementation effort, major risks, and a recommended option.
Output: A comparison table followed by a 150-word recommendation.

2. Provide the right context

Include facts the model cannot infer safely: the audience, purpose, definitions, constraints, prior decisions, and relevant background. Separate instructions from reference material with headings or delimiters so the model can distinguish what to do from what to analyze.

Do not add unrelated context simply because the model accepts a long input. Extra material can obscure the requirement and make it harder to verify the answer.

3. Use examples to show the pattern

One or two representative examples can define a format or classification boundary more precisely than a long description. Show the input and the desired output, then provide the new input. Avoid examples with sensitive data, and make sure they reflect the rule you actually want applied.

Classify each request as Billing, Technical, or Account.

Example:
Input: "I was charged twice."
Output: Billing

New input:
"[customer request]"

4. Set explicit constraints

Specify requirements that matter to the final use: length, tone, reading level, permitted sources, fields, file format, or prohibited content. Positive instructions are usually clearer than a long list of things to avoid.

• “Use plain language and define technical terms” is more actionable than “Do not be confusing.”
• “Return valid JSON with these four keys” is clearer than “Make the answer structured.”
• “Use only the supplied policy” prevents unsupported additions when source fidelity matters.

5. Break complex work into stages

For a large task, request a useful intermediate result before the final output. A research brief might progress from source extraction to comparison, outline, draft, and fact-check. Review each stage instead of letting an early misunderstanding propagate through the entire task.

This does not mean every task needs a lengthy conversation. Use stages when the result is expensive to redo or when intermediate decisions need human judgment.

6. Combine text with relevant files or media

When the model supports the input type, attach the actual image, document, audio, or data file instead of describing it from memory. Tell the model what to inspect and what evidence to return.

Inspect the attached screenshot for layout problems. List each issue with its location, likely user impact, and a suggested fix. Do not infer behavior that is not visible in the image.

For visual generation, provide reference images only when you have permission to use them and describe which characteristics should be followed.

7. Use reusable prompt templates

Templates improve consistency for recurring work. Keep stable instructions in the template and replace only clearly marked variables.

Topic: {topic}
Audience: {audience}
Goal: {goal}
Source material: {source}
Required sections: {sections}
Tone: {tone}
Constraints: {constraints}

Test the template with ordinary, ambiguous, and edge-case inputs. If users often leave a field blank, state how the model should handle missing information.

8. Ground the answer in supplied sources

For factual work, ask the model to base claims on identifiable material and distinguish source facts from inferences. Keywords alone are weak anchors because they do not explain the intended relationship between concepts.

Answer using only the supplied documents. For every key claim, identify the supporting section. If the documents do not contain enough information, state what is missing instead of guessing.

Source grounding makes verification easier, but it does not guarantee accuracy. Check the cited passage and whether it genuinely supports the conclusion.

9. Request a targeted quality check

A generic instruction such as “improve this” gives the model no standard. Name the checks that matter, then ask for a revised version.

Review the draft for unsupported claims, duplicated points, unexplained jargon, and recommendations that do not follow from the evidence. List the problems first, then provide a corrected draft. Preserve any verified quotations and figures.

For high-stakes work, independent human review and authoritative sources remain necessary. A model evaluating its own answer may overlook the same mistake twice.

10. Assign separate roles only when they add value

A multi-step or multi-agent workflow can separate research, drafting, and checking, but extra roles do not automatically improve quality. Define each role's input, output, and review responsibility, and prevent later stages from treating an unverified draft as fact.

• Extractor: Pull relevant facts from approved sources.
• Writer: Create the requested deliverable from those facts.
• Reviewer: Test the draft against a specific checklist.

For many everyday tasks, a single well-structured prompt plus one review pass is simpler and just as effective.

A compact prompt structure

When you are unsure where to start, use five parts: task, context, source material, constraints, and output format. Add an example only when the format or boundary is difficult to explain. After receiving the result, check it against the success criteria and refine the instruction that caused the largest problem.