Roles, Instructions & Examples

Introduction

Module 1 named the parts of a useful prompt once and moved on. This module takes three of them and teaches them properly: the role you give the assistant, the instruction that names the work, and the examples that show what a good result looks like.

Most people arrive believing — reasonably — that better results come from better wording. This module replaces that with something more useful and more durable: results improve when you make your expectations explicit. Wording is the surface. Structure is the substance.

By the end of this module you will be able to look at a disappointing answer and say which single part of your prompt to change — instead of rewriting the whole thing and hoping.

The Three Components

Role + Instruction + Example → an expectation clear enough to judge the result against.

Each answers a different question, and each can be left out — except one.

• •Role — from what point of view, at what depth, for whom? Omit it when the task has no meaningful perspective: arithmetic, formatting, a factual lookup.
• •Instruction — what work should be done? Never omit this. It is the one component a prompt cannot do without.
• •Example — what should the result look like? Omit it when the expected form is obvious, or when you would be guessing at the example yourself.

How this maps onto Module 1: Instruction is the Task part taught properly, and Examples is the Examples part taught properly. A role is not one of the seven — it is a framing choice that sits around the anatomy and often does the work of Audience, Output and Constraints at once. It is not an eighth ingredient every prompt needs. A prompt with a role and an example but no clear instruction is a costume with nobody inside it.

Structure Beats Wording

There are no magic words. Two prompts of almost identical length can produce visibly different results, and the difference is not the phrasing — it is what was made explicit and what was left for the assistant to guess.

This is good news, because it means the skill is learnable and repeatable. You are not hunting for a lucky phrase. You are deciding, deliberately, what the assistant needs to be told.

• •Wording is the surface; structure is the substance.
• •The question is never "how do I phrase this?" but "what have I not said?"
• •Every component you add should remove a guess, not add decoration.

Roles — Choosing a Useful Point of View

A role sets a point of view: which considerations get raised first, how much is assumed and how much explained, whether the vocabulary is plain or technical, who the result is written for, and what "doing this well" means for the task.

Notice what these have in common — each names a job being done, not an identity being adopted.

• •"Act as an editor reviewing this for clarity and structure."
• •"Explain this the way a beginner-friendly science teacher would."
• •"Help me review this plan the way an experienced volunteer coordinator would."

Sometimes the right choice is no role at all. If you cannot say in one sentence what you expect the role to change, the task probably does not need one.

A Role Is a Frame, Not a Credential

A role does not give an AI system real professional credentials, authority, access, or accountability. It changes the shape of the writing. It changes nothing about what the system knows or what it is answerable for.

Compare the framing that gets this wrong with the framing that gets it right:

1. Not: "You are a doctor. Diagnose my symptoms." — Instead: "Using a clinician's structured approach, help me organize clear questions to take to a licensed professional."
2. Not: "You are my lawyer. Is this contract binding?" — Instead: "Help me list the terms in this agreement I should ask a qualified adviser about."
3. Not: "You are a financial adviser. Where should I invest?" — Instead: "Explain, in general terms, how people usually weigh risk against time horizon, so I can prepare questions for a qualified adviser."

The pattern: the role shapes how the help is organized; you still take the decision to a qualified person. In health, law, finance, and public safety this is not a formality — it is the whole point of the boundary. A task role names work: editor, reviewer, organizer, teacher. An immersive persona invites you to treat the system as though it has actually become a real professional with real accountability. Name the job, not the identity.

Instructions — Naming the Work

A good instruction answers one question without ambiguity: what should be done? Depending on the task it may also settle the desired output, the scope, who it is for, any limits, the order of steps where order matters, and — occasionally usefully — what to avoid.

Weak: "Help with my presentation." Nothing is decided. Topic, audience, length, and form are all guesses the assistant now has to make.

Better: "Create a five-point outline for a 10-minute presentation explaining solar energy to high-school students. Use plain language and include one everyday example." The work is named, the audience is named, the form is named, and the length is named.

The second prompt is not better because it is longer. It is better because it can be checked. Five points or not five points. Plain language or not. Everyday example or none. An instruction that cannot be checked cannot be improved, because there is nothing to compare the result to.

• •Instruction — what the assistant should do.
• •Context — the information it needs in order to do it well.
• •"Summarize this for our volunteer committee" is an instruction. The committee's size, what they decided last month, and the fact that two members are new — that is context.

Module 3 teaches context design properly. It is a large subject and it is not this module's — here you only need to tell the two apart.

Examples — Showing What "Good" Looks Like

Some expectations are far quicker to show than to describe. An example can communicate the format a result should take, the tone to write in, the level of detail expected, the structure to follow, the classification logic to apply when sorting things into groups, and the terminology to prefer.

The clearest case is structure. "Write a concise summary" leaves the shape to chance. "Write a concise summary. Use this shape: Problem → Evidence → Recommendation" settles it in nine words.

Working from one or two supplied examples is sometimes called few-shot prompting. It is worth recognizing if you meet the term elsewhere, but it is optional vocabulary, not the lesson.

An example is also an instruction you may not realize you are giving. Before you send one, read it.

• •It can narrow the result too far — one example of a short, formal reply can quietly rule out every other reasonable answer.
• •It can reproduce its own mistakes — an error or an awkward phrase in the example tends to come back in the output.
• •It can carry bias — examples that only ever show one kind of person, place, or situation teach that pattern along with the format.
• •It can encourage copying instead of thinking — a result that imitates the example's content rather than its shape has missed the point.
• •It can go stale — an example built around last year's process keeps producing last year's answer.
• •It can leak private information — the fastest way sensitive data ends up in a prompt is inside a "realistic" example pasted in without a second look.

The example is where private information usually gets pasted, because a real one feels more useful than an invented one. Do not put passwords, keys or tokens, card or account numbers, confidential work information, other people's personal details, or regulated records into a prompt — including inside an example. Make the example synthetic, or change the details until nothing in it identifies anyone. A good example demonstrates a shape; the shape almost never needs the real data.

Putting the Three Together

Build one prompt in four passes on the same task, and watch each component earn its place.

• •Right facts, wrong voice or wrong depth → change the Role.
• •Right voice, but it did the wrong work → change the Instruction.
• •Right work, wrong shape or wrong length → change the Example.
• •It invented details, or got something factually wrong → none of the three. This is a verification problem, not a phrasing problem. Check it yourself.

1. A — "Write something about our new opening hours." You can judge almost nothing; any output is arguably compliant.
2. B — add the instruction: "Write a short notice announcing that we now open at 8am on Saturdays." Now you can judge whether it says the right thing.
3. C — add the role: "Write it the way a friendly community noticeboard would." Now you can judge whether the tone suits the reader.
4. D — add the example: "Use this shape: what changed → when it starts → who to ask." Now you can judge whether the result is usable as it stands.

Longer is not the lesson; checkable is the lesson. A padded prompt — three roles, restated instructions, six examples — is often worse than prompt B, because the assistant now has to guess which parts matter most. Clarity over length: if a component cannot be justified in a sentence, leave it out. When a result disappoints, the loop is RESULT → WHICH COMPONENT? → CHANGE ONE → ASK AGAIN. Change one component at a time; changing three at once produces a different result but teaches nothing. Module 5 teaches evaluating and iterating properly — this is a component locator, not an evaluation method.

Human + AI

You set the purpose, decide what a good result would be, apply judgment and ethical limits, accept or reject the output, and remain responsible for it. The assistant helps generate, organize, restructure, offer alternatives, and iterate.

Roles, instructions and examples are how a person makes their expectations explicit — they are structured communication, not magic words. A role is a framing choice, not a qualification. An example is a demonstration, not a guarantee. You check anything that matters.

• •There is no perfect prompt and no guaranteed output.
• •The assistant does not understand exactly what you meant — it works from what you said.
• •Prompt structure does not eliminate invented detail; it makes it easier to spot.
• •A role does not make the assistant an expert, and examples do not guarantee consistent output.

Writing in Your Strongest Language

Roles, instructions and examples are language-neutral techniques. You can write all three in whichever language you think most precisely in, where the assistant you are using supports it, and the reasoning taught here is unchanged.