← Back to blog

The Role + Example Prompt Pattern That Makes AI Outputs More Consistent

The Role + Example Prompt Pattern That Makes AI Outputs More Consistent

Open an AI chat and ask, “Write this in a professional tone.” The result may be grammatically correct and still miss the audience, the purpose, or the facts that must remain unchanged.

The issue is not that the model does not understand the word “professional”. The issue is that the word does not say who is making the judgement, who will read the output, or what matters most. A reliable pattern is to combine a working role with a representative example.

The role is not theatre. It tells the model which professional lens to use. The example is not decoration or an invitation to copy blindly. It is a format contract: a concrete reference for order, density, labels, and what to do when information is missing.

Give the model a point of view

The same product notes look different to an HR specialist, a support agent, and an editor. HR will look for responsibilities and requirements. Support will look for the next action a customer can take. An editor will look for order and readability. “Make it better” leaves that decision to the model.

Write a role in three layers: professional context, task, and priority. For example:

You are an editor who writes onboarding documents.
Turn the notes below into instructions for new employees.
Prioritise clear steps and preserve limitations. Mark missing information as “to confirm”.

This is more useful than “You are a professional editor” because it describes the job and the boundary for uncertainty.

Use examples as a format contract

If the output needs a repeatable structure, one complete example often beats ten abstract style rules. A good example shows the heading length, field order, sentence density, and the correct treatment of an empty field.

For an event notice, the reference might look like this:

[Event] New employee orientation
[Audience] Employees in their first week
[Time] To confirm
[Preparation] Bring your staff card; use the source notes for other details
[Open items] Venue, owner

Then state: “Keep these fields and this order. When the source does not contain a value, write ‘to confirm’; do not invent one.” The example turns a vague request for a tidy result into something observable.

A three-step pattern

1. Define the role and reader

State the working context, target reader, and priority. Avoid stacking prestige words such as “world-class” or “genius”; they are not acceptance criteria.

2. Provide one or two close examples

The examples should resemble the real task. Include one missing field or one sentence that must remain verbatim. This teaches the boundary, not only the surface tone.

3. Add an acceptance check

Ask for the format used, fields still needing confirmation, and one possible ambiguity after the main answer. The role and example set direction; the check makes uncertainty visible.

Three reusable templates

Writing or rewriting

You are a <role> writing for <reader>.
Rewrite the source below for <purpose>.
Use Example A for tone and Example B for paragraph order and length.
Keep only facts supported by the source. Do not add promises; mark missing data “to confirm”.
After the draft, list three changes and why you made them.

Organising information

You are a traceability-focused <role>.
Organise the material in this format: <paste example>.
Keep a source basis for every item. Put anything you cannot classify under “unclassified”; do not guess.
Report the item count and quote two source lines for a quick check.

Comparing options

You help <reader> make a decision as a <role>.
Compare <Option A> and <Option B> using the format in this example: <paste example>.
List facts first, then recommend against my priority: <priority>.
List missing costs, dates, and constraints separately as “to confirm”.

At this point the method is ready to save. Copy one template, replace the angle-bracket fields, and test it on a small task before making it longer.

Common failure modes

The role is a wish. “Be creative and professional” cannot be checked. “Write onboarding steps for a first-time reader and preserve limitations” can.

The example is from another task. A press-release example will not reliably teach a meeting-notes format. Match the reader, fields, and approximate length.

The example contains unmarked placeholders. A model may treat a sample date or name as a real fact. Label illustrative values clearly.

You check style but not evidence. A convincing voice does not make facts true. Keep a source-basis or “to confirm” field for important documents, and review the final output yourself.

Turn one good result into a reusable method

After a useful run, save the final prompt, one representative output, and the edits you made. Turn repeated edits into rules. Keep only one or two examples that represent the format best. This produces a short, stable template instead of an ever-growing instruction block.

Essevin’s AI chat lets you select from multiple compatible models in one account. You can use a faster model for a first draft and another for a role-and-example format check. The model is a tool; the final judgement remains yours.

Role gives direction, example gives shape, and the acceptance check gives boundaries. Put all three in the same prompt before adding more adjectives.


Information in this article is current as of 12 August 2026 and is provided for general reference only; it does not constitute advice of any kind. Third-party product features, pricing and policies are subject to their official announcements. Essevin service details are as shown on essevin.com and in the console.