How to Write Better AI Prompts With the CREATE Framework

The CREATE framework turns a vague AI request into six useful decisions: Character, Request, Examples, Adjustments, Type of output, and Evaluate. You do not need every element for a simple question, but the checklist helps when the result must follow a specific purpose, audience, format, or source.

The CREATE prompt formula

C — Character: choose a useful perspective

A role can focus terminology and priorities, but it is optional. “Act as an experienced UX researcher” may help with an interview plan; it does not give the model real credentials or guarantee accuracy.

Weak: Write a marketing plan.

Better: Act as a digital marketing manager for a small local service business.

Describe the needed perspective rather than asking for empty authority such as “the world's best expert.” For high-stakes medical, legal, or financial tasks, use AI to organize questions or explain general concepts—not to impersonate a licensed professional.

R — Request: state the task and objective

Use a direct verb: compare, rewrite, extract, classify, draft, debug, explain, or critique. Include the decision or action the answer should support.

Vague: Write about AI in retail.

Specific: Explain three practical ways a five-person online store can use AI to reduce repetitive customer-support work without sending payment data to a chatbot.

A clear task matters more than a complicated persona. If the request contains several independent jobs, number them or handle them in separate turns.

E — Examples or evidence: show what “good” means

Examples are useful for tone, structure, labels, and consistent transformation. Source material is essential when the answer must stay grounded in specific facts.

Example format:
Product: Compact coffee maker
Best for: Small kitchens
Main trade-off: Makes one cup at a time

Use the same three labels for the blender described below.

Tell the model which parts of the example to imitate and which to ignore. One good example is often better than several contradictory ones.

For factual work, paste the approved source and say: Use only the supplied text. If a fact is missing, write “not provided.” This reduces unsupported additions, although the result still needs review.

A — Adjustments: define audience and constraints

Useful constraints include:

• target reader and their current knowledge;
• tone and reading level;
• topics that must be included or excluded;
• length range;
• date, market, platform, or jurisdiction;
• terminology that must remain unchanged;
• how uncertainty and missing information should be handled.

Example: Write for first-time shop owners. Use plain English, define unavoidable technical terms, keep it between 500 and 700 words, and do not invent statistics or customer quotes.

A word limit is a target, not proof of compliance. Count or edit the final text when an exact limit matters.

T — Type of output: specify the deliverable

Ask for the structure you will actually use:

• a table with named columns;
• a numbered procedure with prerequisites;
• valid JSON that follows a supplied schema;
• a short email with a subject line;
• a decision memo with recommendation, evidence, risks, and next steps;
• three alternatives followed by a comparison.

For data formats, provide field names, permitted values, and an example. For prose, name the headings instead of merely saying “make it well structured.”

E — Evaluate: build review into the workflow

No prompt makes an AI response automatically correct. Define checks before relying on it:

• Does every factual claim appear in the supplied source?
• Were all requested fields included?
• Are calculations reproducible?
• Does the output meet the audience and tone requirements?
• Are assumptions labeled?
• Does a human with relevant knowledge approve high-impact decisions?

OpenAI's current prompting guidance similarly emphasizes clear, specific instructions and iterative refinement. The same broad habits help with many chatbots, though tools differ in capabilities, context limits, web access, and safety behavior.

Complete CREATE example

Character:
Act as a UX researcher helping a small mobile-banking team.

Request:
Review the onboarding steps below and identify the five biggest causes of abandonment. Recommend one practical change for each.

Examples or evidence:
Use only the supplied usability notes and screen descriptions. Quote no more than one short phrase per finding. If evidence is missing, label the point as a hypothesis.

Adjustments:
Write for a product manager. Separate observed evidence from inference. Do not invent conversion rates. Prioritize accessibility, trust, and effort required.

Type of output:
Create a table with these columns: Priority, Problem, Evidence, Recommended change, Effort, Validation test.

Evaluate:
After the table, list any recommendation that cannot be supported by the supplied notes and three questions the team should answer next.

A shorter version for everyday work

You can compress the same thinking into one paragraph:

Rewrite the email below for a busy customer who is not technical. Keep the confirmed delivery date and refund terms unchanged, remove internal jargon, use a calm tone, and stay under 180 words. Return a subject line and the email body. Then flag any statement that may need fact-checking.

When CREATE helps most

• rewriting for a defined audience;
• extracting data into a repeatable structure;
• comparing products or options against fixed criteria;
• brainstorming within budget, time, or brand constraints;
• summarizing a supplied document without adding outside claims;
• drafting code, tests, or documentation with explicit requirements.

For exploratory idea generation, TipsMake's brainstorming prompts show how constraints and later evaluation can work together. For summaries, review the limitations discussed in why AI summaries still need checking.

Common prompting mistakes

• Overloading one prompt: Separate unrelated tasks or number the required stages.
• Contradictory instructions: Decide whether brevity, completeness, tone, or a rigid format has priority.
• Examples that conflict with the request: Explain exactly what the model should copy.
• Requesting hidden reasoning: Ask for concise evidence, assumptions, calculations, or a verification checklist instead.
• Sharing confidential data: Remove credentials, personal data, trade secrets, and information you are not authorized to upload.
• Treating fluent text as verified fact: Check primary sources and recalculate important numbers.

Bottom line

CREATE is a checklist, not a magic command. The most important elements are a clear task, enough context, relevant evidence, explicit constraints, and a defined output. The final E—Evaluate—prevents a polished response from being mistaken for a verified one.