Short answer
A useful business prompt for ChatGPT, Claude, Copilot or Gemini has five parts: the task, the context about your business and audience, the material to work from, the rules the output must follow, and the format you want back. Supplying your own facts and telling the assistant to flag gaps instead of inventing them does more for quality than clever wording or asking it to act as an expert.
- Five parts: task, context, material, rules, format.
- Tell the assistant to mark missing facts in brackets rather than invent them.
- Correct a draft with specific feedback instead of starting a new chat.
- The same structure works across ChatGPT, Claude, Copilot and Gemini.
A one-line request to an AI assistant usually gets a bland, generic reply, and the fix is a better brief rather than a better tool. This lesson teaches anyone in a New Zealand business one prompt structure, well enough to use without thinking, and then has you prove it on a task from your own week. You need 20 minutes, any of ChatGPT, Claude, Copilot or Gemini, and one piece of routine writing you do often.
Why most business prompts underperform
A prompt is a brief. A vague brief to a new staff member gets a vague result, and AI is no different: it fills every gap with the most average text it can produce. TRA’s 2025 research for MBIE found many small business users give up after one poor result instead of rephrasing (TRA for MBIE). Usually the brief was the problem.
The five-part structure
| Part | What it answers | Weak version | Strong version |
|---|---|---|---|
| Task | What do you want produced? | “Write an email” | “Write a follow-up email to a customer who has not replied to our quote” |
| Context | Who are we, who is it for, what tone? | Nothing | “Six-person electrical contractor in Dunedin; homeowner customers; friendly, not pushy” |
| Material | What facts must it use? | Nothing, so it invents them | The quote summary, pasted in |
| Rules | What must it do or avoid? | Nothing | “Under 120 words. Do not change any price or date. No discounts.” |
| Format | What shape should the answer take? | Whatever it chooses | “Subject line, email body, then a note listing anything you were unsure of” |
Two rules belong in almost every business prompt:
- “If information is missing, put it in [square brackets] rather than inventing it.” This turns hidden guesses into visible gaps you can fill.
- “Do not add claims about our business that are not in this prompt.” Assistants like to add “fully certified” or “the best service in town”. The Commerce Commission says businesses must have a reasonable basis for claims they make (Commerce Commission), and an AI-written claim is still your claim.
Exercise: rewrite one of your prompts
- Pick a routine writing task. Choose something you write at least weekly: a quote follow-up, a booking confirmation, a supplier query, a reminder to clients.
- Run the lazy version first. Type the one-line request you would normally use. Save the output.
- Write the five-part version. Fill in task, context, material, rules and format. Replace any customer names or contact details with placeholders such as Customer A.
- Run it in a new chat. Save the second output alongside the first.
- Score both out of four. One point each for: facts match your material; tone suits the reader; every rule was followed; you would send it with fewer than three edits.
- Correct, do not restart. Reply to the better draft with specific fixes (“shorter second paragraph; mention the 30-day validity”) and see how close one round of feedback gets you.
Worked example: a Dunedin electrician’s quote follow-up
Illustrative example: the office manager at a six-person electrical contractor in Dunedin follows up quotes that have had no reply after 14 days. Her lazy prompt was “write a follow up email for a quote”. The output promised a “special discount for acting now” and called the business “Dunedin’s most trusted electricians”. Neither was true. Here is her five-part version:
TASK
Write a follow-up email to a homeowner who has not replied to our quote.
CONTEXT
We are a six-person electrical contractor in Dunedin. Most customers are homeowners.
Our tone is friendly and direct, never pushy. We do not offer discounts to chase quotes.
MATERIAL
Customer: Customer A
Quote sent: 14 days ago
Work: switchboard upgrade and six LED downlights in the kitchen
Quote total: $3,480 including GST, valid for 30 days from issue
Earliest start: week of 9 November
RULES
Under 120 words. New Zealand English.
Do not change any price, date or scope. Do not offer a discount.
Do not add claims about our business that are not in this prompt.
If information is missing, put it in [square brackets] rather than inventing it.
FORMAT
Subject line, then the email, then one line headed "Check:" listing anything you were unsure of.
Example output (illustrative and abridged):
Subject: Your switchboard and downlight quote
Hi Customer A,
Just checking in on the quote we sent two weeks ago for your switchboard upgrade
and six LED kitchen downlights ($3,480 incl. GST). The quote is valid for 30 days
from issue, so it is good until [expiry date].
Our earliest start is the week of 9 November. If you would like to go ahead,
or have any questions about the scope, reply to this email or call [phone number].
Cheers,
[Name]
Check: I did not have the exact expiry date, phone number or your name.
It scores four out of four after she fills the brackets. The “Check:” line is the most useful part: it tells her exactly what to verify.
Where prompting goes wrong
- Role-play instead of information. “Act as an expert electrician” does not give the model your prices or the rules you work under. Context and material do; the lesson on giving AI useful context goes deeper.
- Contradictory rules. “Be detailed” and “under 80 words” in the same prompt produce unpredictable results. Read your rules once before running.
- Expecting the prompt to supply facts. No wording makes the model know your stock levels or a client’s history. If it is not in the material, it is a guess.
- Pasting untrusted text. Material copied from emails or websites can contain instructions the assistant may follow. Our lesson on recognising prompt injection explains the risk.
- Treating one run as the answer. Outputs vary between runs and change when vendors update models. Re-test prompts you rely on.
Once a prompt works, keep it. The lesson on building a reusable prompt library shows how to store it with its test input, and our AI prompt library template gives you the format. The ChatGPT and Claude guides explain where each tool lets you save instructions.
Check your prompt
- My prompt has all five parts: task, context, material, rules, format.
- It tells the assistant to bracket missing information rather than invent it.
- It forbids claims about the business that I did not supply.
- Customer names and contact details are replaced with placeholders.
- I scored the lazy and structured outputs and the structured one scored higher.
- I corrected a draft with specific feedback instead of starting over.
Next lesson: Checking AI answers against sources teaches you to verify the facts a well-written draft can still get wrong. Teams who want to practise this together on their own templates can book AI training for business teams.
Sources
- From daunting to doable: positioning AI in reach for SMEs, TRA for MBIE, July 2025
- Making accurate claims, Commerce Commission, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.