Most disappointment with AI comes from prompts like "help with this" or "make it better". The model does its best, which is a guess. A good prompt is not clever; it is specific. Four parts cover almost every case.
1. Goal — what do you want back?
Before: "Look at this proposal." After: "List the three biggest risks in this proposal for the client, in one line each."
2. Context — what should it use?
Point it at the source: "using the attached spreadsheet", "based on the last five emails from Priya", "from the meeting transcript". In Microsoft 365 you can reference files and people directly with a forward slash.
3. Audience and tone
"For a customer who is not technical." "Formal, but warm." "As a two-minute update for the owner." The same content lands completely differently depending on who it is for.
4. Format
"A table with columns for item, owner and date." "Five bullet points." "A single paragraph under 80 words." Say the shape and you will get the shape.
Put together
"From the attached meeting transcript, list the decisions and the action items with owners, as a table, in plain language for someone who missed the meeting." That is a prompt that works every time.
Two habits that raise the ceiling
- Iterate. The first answer is a draft. "Shorter." "More formal." "Add the dates." Each turn improves it.
- Ask it to show its sources. In Microsoft 365 Copilot, answers cite the file or email they came from. Click through before you rely on anything.
We include a short prompting session in every Copilot roll-out we do, because ten minutes of this is the difference between a licence that gets used and one that gets cancelled.
