›

Working with AI

The anatomy of a good prompt

Role, task, context, format, constraints: the five parts of a good prompt.

Listen to this articleAI-generated voice

Most bad AI output is the result of a bad prompt. Most good AI output is the result of a small number of prompting habits that anyone can learn in an afternoon.

A prompt that works has five parts. No jargon. No technique names. Just the five that matter and how to use them.

The five parts

Role, task, context, format, constraints. Not every prompt needs all five: a short job runs on two or three, a complex one usually wants all of them.

  • Role sets the perspective and the level of expertise. "You are a project manager at a 15-person marketing agency" focuses the output. "Act as the world's leading expert in…" rarely does.

  • Task is the only non-negotiable one. Use a verb and be specific. "Help me with my email" gets a generic email; "rewrite this email to be friendlier and shorter" gets a useful one.

  • Context is everything the AI cannot know: who it's for, what happened before, what's fixed, what they've already seen.

  • Format is the part most people forget. Three short paragraphs, a numbered list, a table with three columns: say which.

  • Constraints rule things in or out. Under 200 words. No jargon. Include a clear call to action.

Create in ViraOps is built on this framework: pick a client and a job, type one line, and the request reaches your AI with all five already in place.

A vague prompt

  • "Help me with my email"

  • No role, so it answers like an essay

  • No context, so it invents a situation

  • No format, so you get prose where you wanted a list

  • No limit, so it runs twice as long as it needs to

A specific prompt

  • "Rewrite this email to be friendlier and shorter"

  • "You are a project manager at a 15-person agency"

  • "The client is senior, non-technical, and has seen version one"

  • "Three short paragraphs, no headings"

  • "Under 200 words, no jargon"

Remove the context and the output is technically correct but irrelevant.

Test each part by taking it away

The quickest way to believe any of this is to remove one part and watch what happens. Remove the role and the output reads like an essay instead of a working email. Remove the task and you get something generic. Remove the format and you get a wall of text where you wanted a list. Remove the constraints and it runs twice as long as it needs to, in phrases you'd never use.

Context is the one that fails quietly. The others break visibly. A prompt with no context comes back polished, confident, and about somebody else's project.

Where to start

Five parts is a lot to hold in your head when you're moving fast, and that is exactly when people slip back into vague prompts and generic output. Three of the five do most of the work. Get these right and the other two are refinements.

Three things to do this week

01

State the task in one plain sentence

If you can't say what you want in a line, the AI can't guess it either.

02

Give it the context only you have

Who it's for, what happened before, what's fixed, and what they've already seen.

03

Say what the output should look like

Length, structure, format. Leave it out and you get a wall of text when you wanted a list.

The five-part prompt

ROLE: You are [role] at [type of business]. TASK: [Verb] [what you want]. Write, rewrite, summarise, list, compare, draft, outline, translate. CONTEXT: The audience is [who]. What's happened before: [background]. What's fixed: [constraints]. Already tried: [what]. FORMAT: [Three short paragraphs, no headings / numbered list, each point under 20 words / table with columns: action, owner, deadline]. CONSTRAINTS: Keep it under [N] words. Don't use jargon. [Anything else to avoid or include.] Short tasks may only need Task and Format. Complex ones usually need all five.

Conclusion

01

Five parts: role, task, context, format, constraints. Short jobs need two or three, complex ones need all five.

02

Task, context and format alone will fix most bad prompts.

03

You can test each part by removing it and watching what breaks.

Mark as read

Go to next