R and I: Role and Instructions

Start with an example

Two students ask AI to look at the same essay introduction.

Student A: “Look at my intro.”

Student B:

Role: You are a high school English teacher who grades AP Language essays. Task: Give feedback on the introduction of my argumentative essay. Source: The introduction pasted below. Comment only on what’s there. Action: Identify the three biggest weaknesses and suggest one specific fix for each. Do not rewrite it for me.

Student A gets a vague, cheerful response (“Great start! Consider adding more detail.”) and possibly a full rewrite they didn’t ask for, which they can’t turn in anyway.

Student B gets exactly three useful, specific notes from the perspective of someone who knows what graders look for, and keeps ownership of their own writing.

The difference lives in the first two letters of R.I.C.K.: Role and Instructions.

R — Role: who the AI should act as

AI tools are general-purpose. Left alone, they answer like a smart generalist who knows a little about everything. That’s fine for casual questions. It’s not fine when you need focused, expert-level thinking.

The Role tells the AI which lens to look through. It sets the mindset, the vocabulary, the priorities, and the level of expertise.

Think about it this way: a family doctor and a heart surgeon both know medicine. But if you have a question about your heart, you want the heart surgeon answering. Same knowledge base, very different focus. A Role does the same thing for your prompt.

You don’t need a paragraph. One sentence, sometimes one phrase, is enough:

• “You are an experienced operations manager.”
• “You are a plain-English writing coach.”
• “You are a compliance reviewer familiar with student privacy rules.”
• “You are a marketing copywriter who specializes in small-business social media.”
• “You are a patient, encouraging tutor helping a student who is new to chemistry.”

What changes when you add a Role

Without a Role	With a Role
A general, neutral voice	Vocabulary and focus of that expert
Written for everyone, which means no one in particular	Calibrated to a perspective
Answers only what you asked	Often flags issues an expert would notice, such as a compliance reviewer pointing out a privacy risk you didn’t ask about

A Role is not a license

Telling the AI “You are a doctor” or “You are a lawyer” changes how it sounds, not what it knows or whether it’s right. It’s still predicting text. Anything medical, legal, or financial still needs review by a real, qualified professional.

Think about it: If you ask the same question to “a skeptical investor” and to “an encouraging startup mentor,” you’ll get very different answers. What does that tell you about how much the framing of a question shapes the answer, even with people?

I — Instructions: tell the AI exactly what to do

This is where most people start, and where most people stop. They write the task and call it a prompt. That’s why most prompts underdeliver. The task alone isn’t enough, but it’s still the engine.

Instructions have three parts: Task, Source, and Action.

Task: what needs to be done

The core request. Make it specific and verb-first.

Weak task	Strong task
“Write something about our new policy.”	“Write a two-paragraph announcement introducing our new homework policy to 9th-grade parents.”
“Help me with this email.”	“Rewrite this email to sound more confident and less apologetic, keeping the core message.”
“Summarize this.”	“Summarize the key decisions from these meeting notes in bullet points, one sentence each.”

Notice what changed: the AI now knows the action (write, rewrite, summarize), the scope (this policy, this email, these notes), and often the shape (two paragraphs, bullets).

Source: what material to use

If you’re giving the AI something to work with, such as a document, notes, a draft, data, or talking points, say so, name it, and say what to do with it.

• “Use only the meeting notes I have pasted below.”
• “The source is the policy draft at the end of this prompt. Do not add information that is not in it.”
• “I’ve pasted my original email below. Revise it; do not rewrite it from scratch.”
• “No source provided. Use general knowledge.”

When you don’t name a source, the AI pulls from its general training, which may or may not be what you need. If you have a specific document, reference it. If the AI shouldn’t go beyond it, say so. That one sentence is one of the strongest protections against invented facts.

Action: the specific thing the AI must do

The action is often inside the task, but naming it separately prevents the AI from doing something next to what you needed.

Action verb	What you get
Review	Comments, problems, or risks; the original stays the same
Rewrite	A new version of your text
Summarize	A shorter version with the key points
Extract	Specific items pulled out, like dates or action items
Compare	Similarities and differences
Rank	An ordered list, best to worst
Simplify	The same content at an easier reading level
Explain	Teaching, with reasons

“Review this for compliance risks” and “Rewrite this to be compliant” both sound reasonable. They produce completely different outputs. One hands you a list of problems; the other hands you a new document. Clarity here saves revision time.

The test for strong Instructions

If a capable human assistant received your Instructions with no other information, could they start the work?

If the answer is no, add specifics.

A worked example: one source, three actions

Malik is treasurer of his school’s robotics club. He has a messy page of notes from the club’s budget meeting.

Same Role and Source every time:

Role: You are an organized, detail-oriented club secretary. Source: The meeting notes pasted below. Use only these notes.

Action 1: Extract

Task: Pull out every dollar amount mentioned. Action: Extract each amount with what it was for, in a list. Result: a clean list of 7 expenses he can paste into the budget sheet.

Action 2: Summarize

Task: Summarize the meeting for members who missed it. Action: Summarize in 5 bullets, decisions first. Result: a quick recap he posts in the club group chat.

Action 3: Review

Task: Check whether our planned spending fits the $1,200 budget. Action: Review the amounts and flag anything that would put us over budget. Do not change the plan. Result: a flag that the competition entry fee was counted twice, which saved the club from a $150 mistake.

Same notes, three completely different, useful results. The Action decided what came back.

Think deeper: the “smart generalist” problem

Why does a Role work at all? Because of how these models learned.

A language model learned from huge amounts of writing by all kinds of people: engineers, poets, lawyers, teenagers, chefs. When you ask a question with no Role, it effectively blends all of those voices into one average answer. When you name a Role, you’re saying “answer like this group of people would,” and the model’s predictions shift toward the words, priorities, and habits it saw in that kind of writing.

The Source line works for a related reason. Without a source, the model answers from its blurry, general memory of the world, which is where invented facts come from. With a source, it has something concrete in front of it to work from.

Questions to sit with:

• If a Role can make the same AI sound like a supportive coach or a harsh critic, how should you decide which perspective you actually need, instead of which one feels best to hear?
• When might it be smart to ask for two roles on purpose, such as “first respond as a supporter, then as a skeptic”?

Practice: one task, three roles

1. Choose a real task, such as feedback on a paragraph, a plan for an event, or an explanation of a hard topic.
2. Write strong Instructions: Task, Source, and Action.
3. Run it three times, changing only the Role: no Role, a general Role, and a specific Role.
4. Compare the three answers.
5. List two differences the Role alone created.
6. Reflect: Which answer was most useful, and was it the same one that felt most pleasant to read?

ROLE:
You are [specific role, expertise, or perspective].

INSTRUCTIONS:
- Task: [verb first: what needs to be done, with scope and format]
- Source: [the material to use, e.g. 'the notes pasted below', or 'no source provided']
- Action: [the exact action: write, rewrite, summarize, compare, extract, rank, review, simplify, explain]