Joel BrilliantI build with AI. Here's how I do it.
itsbrill.ai
Brisbane, Australia

Chapter 2 of 8

Learn the power-user loop

Treat each prompt like a short work brief. Tell Copilot what the job is, give it the right material, then check the result before it leaves your screen.

8 min readOutcome: one reusable methodFacts checked 19 July 2026

01Give every job five parts

The biggest jump in Copilot skill isn't learning a secret phrase. It's learning to hand over work clearly. Microsoft teaches four prompt ingredients: goal, context, source and expectations. This guide adds a fifth because the answer still needs judgement.

The power-user loop

1. GoalWhat job needs doing?
2. ContextWho is this for, and what matters?
3. SourceWhat material can it rely on?
4. ExpectationsWhat should the answer look and sound like?
5. CheckHow will you test the result?

You won't need five labelled paragraphs for every small task. You do need the five decisions. Once they're obvious to you, your prompts get shorter and your results get easier to trust.

Source note. Microsoft's current prompting guidance uses goal, context, source and expectations, and recommends iterating and checking responses. See Get started writing prompts in Microsoft 365 Copilot.

02See the difference in one ordinary job

Imagine you're writing a project update for a senior manager. The source material is a page of fictional meeting notes.

Thin handover

Write a project update from these notes.

Useful handover

Draft a project update for a time-poor senior manager. Use only the attached notes. Lead with status, then decisions needed, risks and next actions. Keep it under 180 words and use plain Australian English. After the draft, list any claim you couldn't support from the notes.

The second prompt gives Copilot fewer places to guess. It also gives you a clear test: audience, order, length, tone and unsupported claims.

Reusable task brief
Goal: [the job you want done]

Context: [audience, situation and important constraints]

Source: Use only [the file, notes, email thread or text provided].

Expectations: [length, structure, tone and format]

Check: Before you finish, [the checks Copilot should run]. Flag anything the source doesn't support.

03Give it a source boundary

Copilot can be fluent about the wrong thing. A source boundary tells it where the answer is allowed to come from.

  • For a summary: use only the attached document.
  • For a status update: use the meeting notes and action list below.
  • For research: cite every factual claim with a link and separate facts from your inference.
  • For work-based chat: name the exact meeting, file or email thread where you can.

“Use our project files” is a large search area. “Use the Project Juniper decision log and the 18 July steering meeting” is a usable boundary.

Before you add a source

Use material your organisation allows in Copilot. Access to a file doesn't automatically mean it belongs in this task or should appear in the output.

04Check four things before you use the answer

Copilot's answer is a draft with a head start. Review it for:

  1. Accuracy. Do names, dates, numbers and claims match the source?
  2. Coverage. Did it include the decisions, risks or actions that mattered?
  3. Audience. Would the intended reader understand what happens next?
  4. Tone. Does it sound appropriate for the person and situation?
Ask Copilot to expose weak spots
Review your draft against the source and the brief.

Return:
1. any factual claim that isn't directly supported
2. any important source detail you left out
3. any wording that sounds more certain than the source
4. the revised draft

Don't invent the missing information. Mark it as a question for me.

This second pass helps you inspect the draft. It doesn't replace your own check against the source.

Source note. Microsoft recommends reviewing Copilot output for clarity, accuracy, tone, audience fit and coverage. See Evaluate Copilot output.

05Save the brief when the job repeats

When a prompt works twice, stop rebuilding it from memory. Save the structure, remove the task-specific material and give it a clear name such as “Weekly project update from meeting notes”.

Keep the stable instructions in the saved prompt. Add the changing source and context when you run it. That separation is what turns a good chat into a repeatable way of working.

Chapter complete

You have the loop

Next, you'll collect the useful context from ChatGPT without dragging private chat history or confidential material into work.