How to write AI prompts for real work

A good prompt briefs the AI the way you would brief a colleague: who it should act as, what you want, the facts it needs, and the shape of the answer. Set those four, then iterate. This guide gives you the structure, the patterns worth reusing, and copy-paste templates for real work tasks.

What prompt engineering actually means

It is a grand term for a simple habit: giving clear instructions so you get a useful result first time instead of a vague one you have to fix. If you have ever handed a task to someone and watched them guess at half of it because you were rushed, you already know the failure mode. The clearer the brief, the less rework you get back.

You do not need to learn this to use ChatGPT, Claude or Copilot. You need it to stop wasting twenty minutes nudging a tool toward something you could have asked for in one go.

The five parts of a prompt that works

That last line matters. It turns a confident draft into one that tells you where it filled gaps, which is where these tools quietly go wrong.

Role. Tell the AI who to act as and who the output is for. "You are a delivery manager writing for executives" sets the tone and the assumptions in one line.
Goal. Say what you want out: a plan, a summary, a table, a first draft.
Context. Give it the facts and the limits. The audience, the constraints, what already exists, what it must not include.
Format. Specify the shape: bullets, a table, a script, a word count.
Iterate. Ask for two or three versions, pick a direction, then refine. Treat the first answer as a starting point.

Work example

You are a delivery manager. Turn these messy meeting notes into a RAID log.
Context: executive audience, plain English, UK spelling.
Format: a table with Risk, Assumption, Issue, Dependency columns, plus three
actions for next sprint.
Then flag anything in the notes you had to guess.

Home Example

You are a household planner. Build a Monday to Friday plan covering school runs,
two training evenings and meal prep for a family of five.
Format: a table with time, task and who.
Add a fallback row for when training is cancelled.

Patterns worth reusing

Once the five parts are second nature, a few patterns do most of the heavy lifting:

Combine them. Persona plus fact-check is strong for anything that has to be both readable and trustworthy.

Persona. "You are a [role] writing for [audience]." Sets voice and assumptions in one move.
Plan. "Give me a step by step for [task], with realistic constraints." Good for anything you will hand to someone else.
Fact-check. "List any claims in your answer that need verifying, and how to verify them." The single most useful line for work output.
Memory. "Hold these standing facts for this chat: [brand tone, house rules, project details]. Confirm you have them." Saves repeating yourself.
Reflection. "Before you finish, check against this list: UK spelling, under 150 words, no jargon." Catches the obvious misses.

From vague to usable in about 90 seconds

The gap between a poor prompt and a good one is rarely effort. It is specifics.

Vague:

Write a business case.

Better:

You are a programme manager. Write a one-page business case for adopting [tool]. Audience: finance and IT. Headings: problem, options, recommendation, cost, risks. Under 400 words. Then give three risks, each with one mitigation.

The second version gives you something you can put in front of people. The first gives you a wall of filler you then have to rewrite.

Common mistakes and how to fix them

Vague ask, vague answer. Fix: state the role and the outcome in one line before anything else.
A wall of text comes back. Fix: force a format. A table or a set of bullets stops it rambling.
One and done. Fix: ask for two or three versions, choose a direction, then iterate.
You trust it blind. Fix: ask it to flag what it guessed or what needs checking, then check those points yourself.

Copy-paste templates

Work first, with one for home. Swap the bits in brackets.

Meeting notes to RAID log:

Weekly status for execs:

Template: Meeting notes to RAID log

You are a delivery manager. Turn these notes into a RAID log for an executive
audience. Format: a Markdown table with Risk, Assumption, Issue, Dependency
columns, plus three actions for next week. Flag anything you had to infer.
Notes: [paste]

One-page business case:

Template: Weekly status for execs

You are a delivery manager. Summarise this week in 150 words: wins, risks,
decisions needed. Audience: executives. Style: calm, factual, UK spelling.
Add three bullets for next week. Source: [paste]

Weekly household plan:

Template: One-page business case

You are a programme manager. Write a one-page business case for [tool/change].
Audience: finance and IT. Headings: problem, options, recommendation, cost,
risks. Under 400 words. End with three risks and a mitigation for each.

Template: Weekly household plan

You are a household planner. Build a Monday to Friday plan around: training
[times], school runs, a shop run and meal prep. Output a table with time, task,
who and kit needed, plus five "if it rains" alternatives.

FAQ

What is prompt engineering?

Giving the AI clear, structured instructions so you get a useful result first time. Role, goal, context, format, then iterate.

Do I need it to use ChatGPT or Copilot?

No. But a simple structure gets you a usable answer faster and cuts the back and forth.

What is the single biggest fix?

Add a role and a format. "You are a [role]" plus "give it to me as a [table/bullets]" fixes most weak prompts on its own.

How do I stop the AI making things up?

Ask it to list any claims that need verifying and how to verify them, then check those points before you use the output. Never ship an AI draft you have not read against the facts.