M365 Copilot — Prompt Hygiene · Quick Start Guide

At a glance

The problem
Vague prompts produce vague answers.
Copilot has no way to ask you to clarify.

The fix
Four components: Role, Task, Context, Format.
Each one removes a guess Copilot has to make.

The payoff
Answers that are specific, scoped, and usable
without a second round of prompting.

The four components — what each one does

Role

Tell Copilot what lens to use. Sets the expertise level, tone, and perspective for the answer.

Act as a project controller
Write this for a non-technical audience
Review this as a procurement manager would

Task

Start with a verb. Copilot needs a clear action to take — not a topic to discuss.

Summarise / Draft / Extract / Compare
Rewrite / List / Identify / Translate
Flag / Score / Group / Explain

Context

Supply what Copilot cannot infer: the specific source, the audience, the constraints, the deadline. Reference files with @.

from @Risk-Register.xlsx
for the ExCom, who haven't seen the data
focus only on overdue items

Format

Specify how the output should look. Without this, Copilot chooses — and often chooses wrong.

in a table with columns: owner, status, due date
in under 150 words
as three bullet points, no preamble

The formula — copy this pattern

Prompt structure

[Task verb] + [What / Source] + [For whom / Constraints] + [Output format]

Example: "Summarise the open risks in @Q2-Risk-Register.xlsx for the project director, who needs a one-page brief before the gate review. List by severity. Exclude risks already closed. Under 200 words."

Before and after — the gap in practice

Summarising a report

Weak prompt

"Summarise this report."

Length and focus are Copilot's choice. May miss the section that matters. Hard to reuse.

Strong prompt

"Summarise @Q1-Performance-Report.pdf for the CFO in 5 bullet points. Focus on financial performance and delivery schedule variance only. No preamble."

Scoped to source, audience, length, and topic. Output is ready to forward.

Components used: Task · Context (@file) · Audience · Format

Drafting a follow-up email

Weak prompt

"Write an email to the client about the delay."

Copilot invents tone, length, and content. May over-commit to a new date or sound apologetic in the wrong register.

Strong prompt

"Draft a 3-sentence email to the client acknowledging the 2-week delay in the piping deliverable. Tone: professional, not defensive. Do not offer a new delivery date. Do not use the word 'unfortunately'."

Tone, length, and constraints are explicit. Negative instructions prevent the most common failures.

Components used: Task · Context (situation) · Constraints · Format + negative instructions

Extracting action items

Weak prompt

"What were the action items from yesterday's meeting?"

Copilot searches broadly. May return items from the wrong meeting or miss informal commitments.

Strong prompt

"Extract every action item from @Project sync Apr 9. For each: owner, description, due date if stated. Present as a table. Flag any item with no clear owner."

Sourced from one specific meeting. Structured output. Gaps are visible (no owner = flagged).

Components used: Task · Context (@meeting) · Format (table) · Edge case instruction

Patterns that fail

The most common weak prompts — and what to write instead

Fails

"Make it better."

Copilot has no definition of "better." It will change something — often the wrong thing.

Works

"Rewrite this to be more direct. Shorten each paragraph by half. Keep all numbers unchanged."

Specific action, specific constraint, specific exception.

Fails

"What are the main risks?"

No source. Copilot guesses which project, which period, which format you want.

Works

"List the top 5 risks in @Risk-Register.xlsx with a score above 12. Table format: risk, owner, score, status."

Source, filter, format — nothing to guess.

Fails

"Summarise my emails and tell me what to prioritise and draft replies and flag anything urgent."

Four tasks in one prompt. Copilot will attempt all four partially.

Works

"Summarise my unread emails from today. List by sender and subject only. No body text."

One task. Run a second prompt for the reply draft once you see what needs one.

Fails

"Analyse this data."

No source referenced. No analysis type specified. Copilot produces a generic description.

Works

"In @Budget-Apr.xlsx, identify any line items where actual spend exceeds budget by more than 10%. List them with the variance amount."

Source, metric, threshold, output — fully specified.

Seven rules

1Start every prompt with a verb. Summarise, draft, extract, compare, rewrite, list, flag, translate. No verb = no task.
2Reference the source explicitly with @. Without a reference, Copilot searches broadly and may use the wrong content. See the Context IQ guide.
3Name the audience. CFO, new joiner, external client, project team — the audience determines the level of detail and assumed knowledge.
4Set a format constraint. Table, bullet list, under 150 words, numbered steps. Without a format, Copilot decides — and often chooses too long.
5Add one negative instruction for precision. "Do not include internal costs." "Exclude items already closed." "Do not offer a new deadline." Negative instructions prevent the most predictable errors.
6One task per prompt. If you need a summary and a draft reply, run two prompts. Multi-task prompts produce partial results on each task.
7Prompt in your working language — Copilot follows the prompt. Copilot generates output in the language you write the prompt in, regardless of the document's language. If you want French output from an English document, write the prompt in French. If the output language keeps switching, add "Respond in [language]" as a final instruction.