A prompt that works isn't clever wording — it's context. Give the AI five things: who it's acting as, what the task really is, the raw material, the format you want back, and an example of what "good" looks like. That takes about ninety seconds to write, and it's the whole difference between generic filler and something you'd actually send.
Nearly every "my AI is useless" conversation I have ends the same way. The tool is fine. The prompt was three words long, with none of the information a human would have needed either. Here's the fix — no magic phrases, no swipe file.
What actually makes a prompt work?
The single most useful mental shift: stop typing requests, start writing briefs.
If a new staff member started on Monday and you asked them to "write a follow-up email to a client", they'd come back with questions. Which client? What happened last time? How formal are we? You'd never judge them on the version they wrote without asking — but that's exactly the position you put AI in when you type one line and hit enter. It can't ask, so it guesses average, because average is what it was trained on.
So what you get back is only as good as the brief you gave. Not the phrasing — the context.
What are the five parts of a good prompt?
Every prompt I write for real work has the same five parts. You won't always need all five, but when something comes back wrong, the missing part is nearly always on this list.
| Part | What it does | What it looks like |
|---|---|---|
| 1. Role | Sets the vocabulary and the level of detail | "You're helping the owner of a 6-person plumbing business." |
| 2. Task | Says what you actually want, and who it's for | "Draft a reply to this client chasing an overdue quote." |
| 3. Raw material | Gives it the facts so it doesn't invent them | The actual email thread, the actual numbers, pasted in. |
| 4. Format | Removes the reformatting you'd otherwise do by hand | "Under 120 words. No bullet points. One clear ask at the end." |
| 5. A sample of good | Teaches your tone faster than describing it ever will | "Match the voice of this email I sent last month: [paste]." |
Part five is the one nearly everyone skips, and it's the one that changes the most. Describing your tone ("professional but warm, not too corporate") is vague. Pasting two emails you actually wrote is precise.
What does this look like on a real task?
Here's the same job, done badly and done properly.
The version that doesn't work:
"Write a follow-up email to a client about an overdue invoice."
What comes back is technically an email. It's also stiff, twice as long as it should be, contains a line about "valued partnership", and needs so much rewriting you may as well have started from scratch. This is where most people decide AI isn't for them.
The version that works:
"You're helping me run a small commercial cleaning business in Melbourne. Draft a follow-up email to a client whose invoice is 19 days overdue. We've worked with them for three years and I want to keep the relationship — this is a nudge, not a demand. Keep it under 100 words, plain and friendly, no corporate language, and finish with one clear question. Here's the original invoice email I sent [paste], and here's an email I wrote to another client last month that sounds exactly how I want to sound [paste]."
That's about forty seconds of typing, most of it pasting — and the result is usually sendable with one small edit.
The rule of thumb: if the answer came back generic, the prompt was generic. Add the part you left out — usually the raw material or the sample of good — rather than rewording the request.
Why is the first answer always a bit off?
Because you're not writing a prompt, you're having a conversation — and almost nobody treats it that way.
The first response is a draft. Your job is to do what you'd do with a junior's draft: say what's wrong with it. "Too formal, cut it by a third." "You've invented a discount we never offered — use only the numbers in the thread." "The second paragraph is the good one, build the whole thing around that."
Two or three rounds of that gets you somewhere a perfect opening prompt rarely does. Iterating isn't a sign you prompted badly — it's the actual skill.
One habit worth building early: when the edits finally produce something you'd send, ask the tool to write the prompt that would have produced this first time — and save that. You've just turned a ten-minute fiddle into a fifteen-second task, permanently.
Which prompts are worth saving?
This is where the real time comes back, and it's the part people miss. Prompt skill isn't about writing better one-offs — it's about building a small library for the things you do every single week. Almost every business I work with has six to ten hiding in plain sight:
- The weekly summary you rebuild from the same three reports.
- The quote follow-up you send four times a week with the details swapped.
- The new-client welcome email and onboarding checklist.
- The meeting notes you turn into actions for someone else.
- The monthly client update that always takes an hour longer than it should.
Write each of those once, properly, with all five parts. Keep them in a doc, a notes app, anywhere you can copy from. Leave obvious blanks — [client name], [paste invoice], [paste notes] — and fill them in each time.
That's the same principle behind every result I've had with clients. The reporting job that went from 20+ hours a week to under 3 wasn't one clever prompt — it was a weekly process, mapped once and then automated so it runs the same way every time. Client onboarding (15 hours to 3) and invoicing (5 hours a week to about 30 minutes) followed the same pattern: find the task that repeats, tidy the process, then make the repeat automatic. A saved prompt is the smallest version of that idea. Saved and repeated beats brilliant and one-off, every time. If you want the wider version of that argument, it's in why most AI projects fail.
What should you never put in a prompt?
Context is the whole game, so the temptation is to paste everything. Don't — decide your rule before you start, not after.
The short version: a paid business or team tier with training switched off is a reasonable place for ordinary internal work. A free consumer account is not the place for client records, contracts, payroll or anything covered by privacy obligations. Redact names and account numbers when the AI doesn't need them — it usually doesn't. The full plain-English guide is in is ChatGPT safe for business data?, and it's worth ten minutes before you roll this out to a team.
What's the fastest way to get better at this?
Pick one task you do weekly. Write the five-part prompt for it. Use it every week for a month, improving it slightly each time.
That's it. You'll learn more from one prompt used twenty times than from twenty prompts used once — and you'll have something permanently faster at the end of it. If you're starting from nothing, how to actually use AI in a small business covers how to choose that first task.
The short version
- Write briefs, not requests — the context is the prompt.
- Five parts: role, task, raw material, format, and an example of good.
- Pasting a sample of your own writing beats describing your tone.
- The first answer is a draft. Edit it in conversation, like a junior's work.
- Save the prompts for weekly tasks — repetition is where the hours come back.
- Decide your data rule before you paste anything real.
So where do you start tomorrow?
Open the last thing you wrote from scratch that you've written a version of before. Write the five-part prompt for it while it's fresh, edit it in conversation until it's genuinely sendable, then save it.
You've now got one task permanently faster, and the habit that makes the next one quicker. Do that six times and your week looks different. If you'd rather build that library with someone in the room, that's what a session with me looks like — details on the services page.
Frequently asked questions
Do I need to learn prompt engineering?
No. Prompt engineering as a discipline is for people building software on top of AI models. For business tasks you need something much simpler: the habit of giving the tool the same context you would give a new staff member. Who it is acting as, what the task really is, the raw material, the format you want back, and an example of good. That habit takes an afternoon to learn and a few weeks of real use to become automatic.
Why does AI keep giving me generic answers?
Because it is working from a generic request. If the only thing you supply is the topic, the only thing it can produce is the average of everything ever written on that topic. Specificity is the fix: paste the actual email, the actual numbers, the actual client situation, and say who the answer is for. The moment the input stops being generic, the output does too.
How long should a prompt be?
Long enough to carry the context, and no longer. In practice most useful business prompts run three to ten lines plus whatever raw material you paste in. Length is not the point though. A ten-line prompt that never says who the reader is will still miss, while a three-line prompt with the real email pasted underneath often lands first time.
Should I use prompt templates I find online?
Use them as a starting shape, not a script. The ones you find online are written for a generic business, so they carry none of the context that makes an answer usable in yours. Take the structure, strip out the theatrical language, and replace the generic parts with your own tone, your own clients and your own examples of good. Then save your version, because that one is worth keeping.
Is it safe to paste real business information into a prompt?
It depends entirely on which account you are using. A business or team tier with training switched off is a reasonable place for ordinary internal work, while a free consumer account is not the place for client records, contracts, payroll or anything covered by privacy obligations. Decide the rule once, write it down, and make sure everyone follows the same one before you start pasting.