Prompt Lab13 min read

How to use ChatGPT effectively: the 5-part briefing prompt (real before and after)

It's 7 p.m., a supplier just pushed back a delivery, and you ask ChatGPT to write the email to your client. What comes back is polite, long and generic, with blanks in brackets where your facts should be. The AI didn't fail: nobody told it who you are, what happened or what the email has to say. Here's the fix we tested for real: a five-part brief for ChatGPT, Claude, Gemini or Copilot, the one line that makes AI ask before it guesses, and the real answers, before and after.

The video loads from YouTube (Google) only if you click. Watch on YouTube

Why ChatGPT gives you generic answers

An AI assistant knows a lot about the world and nothing about your Tuesday. When a request leaves out who you are, who will read the result and what actually happened, the model has two options: guess, or hand you a template full of blanks. Anthropic's own guide for its Claude models puts it plainly:

Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.

Anthropic, Claude Platform Docs: Prompting best practices (consulted Oct. 11, 2026)

The four companies behind the assistants most small businesses use all publish prompting advice. Read side by side, it comes down to the same thing: give the job, the context and the shape of the answer.

What the AI companies' own guides say (consulted October 11, 2026)
CompanyGuideWhat it says
OpenAIHelp Center: best practices for prompt engineering"Be specific, descriptive and as detailed as possible about the desired context, outcome, length, format, style, etc"
GoogleGemini help for Docs, Sheets, Slides, Vids & FormsFour areas: Persona, Task, Context, Format. "You don't need to use all 4, but using a few will help."
MicrosoftCopilot support (last updated September 2026)Four parts: Goal, Context, Source, Expectations. "You can put a little or a lot into a prompt, but all that's required is a clear goal."
AnthropicClaude docs: prompting best practices"Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too."

The test: one line versus a five-part brief

We ran the same request twice, for real, each time in a fresh conversation with no other context. The answers below come from Claude, Anthropic's model, on October 11, 2026. The method is the same in ChatGPT, Gemini or Copilot; only the wording of the answers would change. First, the one-liner most of us type:

PromptPrompt A: the one-liner
Write an email to a client telling them their job is delayed.

What came back was a template: 346 words in all, with 14 blanks in square brackets in the email, from [Project Name] to [Phone / Email], then a list of tips and an offer to tailor it if we told it the job, the delay and the cause. That's a reasonable answer to a request with no facts in it. It's just not an email you can send.

Real AI answerReal answer A (excerpt: the first lines, verbatim)
Here's a professional, adaptable email. Fill in the bracketed parts:

---

**Subject:** Update on [Project Name]: revised completion date

Hi [Client Name],

I'm writing about [project/job name]. Unfortunately, we won't be able to deliver by [original date] as planned. Our new completion date is **[new date]**.

The 5-part briefing prompt

Brief the AI the way you'd brief a new employee on their first morning. Five short parts, in this order:

  1. WHO

    Who you are and who will read the result. One line: business, size, city, audience.

  2. TASK

    One verb, one deliverable. "Write an email telling the client their installation is delayed", not "help me with a client".

  3. FACTS

    Paste the facts and add: use only these, add nothing. This is exactly what the one-liner was missing.

  4. RULES

    Tone, language, length, and what to do instead of guessing. Say what to do, not only what to avoid.

  5. FORMAT

    The shape of the answer: a subject line, then the email, ending with one clear next step.

PromptThe briefing prompt (copy it, fill in the [BRACKETS])
WHO: I'm [WHO YOU ARE: role, business, city]. This is for [WHO WILL READ IT].
TASK: [ONE VERB + ONE DELIVERABLE, e.g. "Write an email telling a client their installation is delayed"].
FACTS (use only these, add nothing):
- [FACT 1]
- [FACT 2]
- [FACT 3]
RULES: [TONE], [LANGUAGE], under [N] words. Don't promise anything that isn't in the facts.
FORMAT: [THE SHAPE, e.g. "A subject line, then the email. End with one clear next step."]
Before you write anything: if something important is missing, ask me up to three questions. Don't guess.

Here's the same delay, briefed. The company is fictional: a 12-person plumbing firm in Laval whose supplier pushed back a water heater.

PromptPrompt B: the same task, briefed
WHO: I run a 12-person plumbing company in Laval, Quebec. This email goes to a residential client.
TASK: Write an email telling the client their water heater installation is delayed.
FACTS (use only these, add nothing):
- Client: Karine.
- The installation was booked for Thursday, October 15.
- Our supplier delayed the 60-gallon water heater. It now arrives Monday, October 19.
- New proposed slot: Tuesday, October 20, 8 a.m. to noon.
- Her current water heater still works until then.
- We're taking $50 off the labour for the trouble.
- She can confirm the new slot by replying, or call 450-555-0142.
RULES: Warm and plain, no corporate apologies. Under 120 words. Don't promise anything that isn't in the facts.
FORMAT: A subject line, then the email. End with one clear next step. Sign it "Marc".
Real AI answerReal answer B (verbatim)
Subject: Your water heater installation: new date, Tuesday, October 20

Hi Karine,

Our supplier delayed your 60-gallon water heater, so we can't install it on Thursday, October 15 as planned. It now arrives Monday, October 19.

We'd like to come Tuesday, October 20, between 8 a.m. and noon. Your current water heater still works until then.

We're taking $50 off the labour for the trouble.

To confirm Tuesday, just reply to this email or call us at 450-555-0142.

Thanks,
Marc

Seventy words in the body. Every fact is there, nothing was added, and it ends on the one thing Karine has to do. You still read it before you send it, but there's nothing left to fill in.

The line that makes AI ask before it guesses

Sometimes you don't have all the facts yet. That's when one line at the end of the brief does the most work. We took the new date, the discount and the phone number out of the brief and added this:

PromptThe ask-first line
Before you write anything: if something important is missing, ask me up to three questions. Don't guess.
Real AI answerReal answer (verbatim)
Before I write it, I have three questions. Each answer changes what the email can say.

1. **Is there a new date?** Do you have a new installation date, or at least a rough window? If not, what should Karine expect next? For example, you'll call her once the supplier confirms. Right now the email can only tell her the 15th is off.
2. **French or English?** Should this go to her in French or in English?
3. **Does she have hot water right now?** Is her current water heater still working? If it isn't, the email should say so directly. If you're doing anything for her in the meantime, tell me and I'll include it. Otherwise I'll leave it out.

If you'd like a phone number or other contact detail in the email, send me that too.

No email, three questions, and one we hadn't thought of: French or English? For a business in Quebec, that's the question that matters. Answer them, and the next reply is the email.

It works for personal tasks too: a double charge on your bill

The brief isn't only for work. Here's an everyday case: your internet bill was charged twice. Notice the rule that keeps your account number out of the conversation. The AI leaves a placeholder, and you fill it in yourself.

PromptPrompt: a refund for a double charge
WHO: I'm a customer of an internet provider. This message goes to their billing support chat.
TASK: Write a message asking for a refund of a double charge.
FACTS (use only these, add nothing):
- My October bill of $84.50 was charged twice to my credit card on October 3.
- I only have one internet account with them.
- I want the extra $84.50 refunded to the same card, not as a credit on my account.
RULES: Polite but firm. Under 80 words. Ask for a reference number for my request. Leave [ACCOUNT NUMBER] and [MY NAME] as placeholders: I'll fill them in myself.
FORMAT: The message only, ready to paste.
Real AI answerReal answer (verbatim)
Hello,

My October bill of $84.50 was charged twice to my credit card on October 3. I only have one internet account with you (account [ACCOUNT NUMBER]), so one of these charges is a duplicate.

Please refund the extra $84.50 to the same credit card. I do not want it applied as a credit on my account.

Could you also give me a reference number for this request?

Thank you,
[MY NAME]

Seventy-two words, firm without being rude, a reference number requested, and neither your account number nor your name went into the chat.

Show it one example of how you write

Describing a tone is hard. Showing it is easy. OpenAI's guide recommends articulating the desired output format through examples. Paste one message you've already sent and ask the AI to match its tone and length:

PromptPrompt: an overdue-invoice reminder in your own voice
WHO: I run a 12-person plumbing company in Laval, Quebec. This email goes to a residential client.
TASK: Write a friendly reminder that an invoice is overdue.
FACTS (use only these, add nothing):
- Client: Louis.
- Invoice 2031, $1,240, bathroom repair, was due September 30.
- He can pay by Interac e-Transfer to billing@example.com or by cheque.
RULES: Match the tone and length of my example below. No threats, no late fees.
FORMAT: The email only. Sign it "Marc".
MY EXAMPLE (an email I sent last month):
"Hi Louis, quick update: the part came in, so we're on for Wednesday at 9. If anything changes, I'll text you the night before. Thanks for your patience! — Marc"
Real AI answerReal answer (verbatim)
Hi Louis, quick reminder: invoice 2031 for the bathroom repair ($1,240) was due September 30 and is still open. You can pay by Interac e-Transfer to billing@example.com or by cheque, whichever is easier for you. Thanks! — Marc

Thirty-eight words, the same rhythm as the example, the same sign-off, no threats and no late fees.

Ask for Québec French, and name your words

If your clients write to you in French, put the language in the rules, precisely: Québec French, vouvoiement or tutoiement, and the words you use. We briefed the Laval email again in French, with one rule: "courriel", never "e-mail" or "mail". Here's what came back:

Real AI answerReal answer in French (verbatim)
Objet : Nouvelle date pour l'installation de votre chauffe-eau

Bonjour Karine,

Petit contretemps : notre fournisseur a retardé votre chauffe-eau de 60 gallons. Il arrive maintenant le lundi 19 octobre, donc on ne pourra pas faire l'installation le jeudi 15 octobre comme prévu. Désolé pour le changement.

On vous propose le mardi 20 octobre, de 8 h à midi. D'ici là, votre chauffe-eau actuel fonctionne encore.

Pour le dérangement, on retire 50 $ de la main-d'œuvre.

Est-ce que cette nouvelle plage vous convient? Pour la confirmer, répondez simplement à ce courriel ou appelez-nous au 450 555-0142.

Merci,
Marc

Natural, warm, vouvoiement throughout, and "courriel" exactly where we asked for it. The full French prompt is in the French version of this article.

Make it check its own draft

Before anything leaves your hands, paste the draft back with your rules and ask for a check. We gave the AI the one-line template from the start of this test, with four rules: only our facts, under 120 words, no brackets, and one clear next step at the end.

PromptThe self-check prompt (copy it, fill in the [BRACKETS])
Check the draft below against my rules. List every rule it breaks, one short line each. Then rewrite it so it follows every rule.

MY RULES:
1. Use only these facts: [YOUR FACTS].
2. Under [N] words.
3. No placeholders in brackets.
4. End with one clear next step.

THE DRAFT:
[PASTE THE DRAFT]
Real AI answerReal answer (excerpt: the rules it flagged, verbatim)
**Rules the draft breaks**

1. **Rule 1 (only your facts):** It makes up content you never gave: a "completion date," sample reasons, fixes, an unchanged price, a partial delivery and a promised follow-up update. It also leaves out every fact you did give: October 15, the water heater, October 20 from 8 a.m. to noon, and the $50 off.
2. **Rule 2 (under 120 words):** It runs to about 200 words.
3. **Rule 3 (no bracketed placeholders):** It has 14 of them, from [Project Name] to [Phone / Email].
4. **Rule 4 (end with one clear next step):** It ends with a thank-you and a signature block. The only next steps in it are vague and split in two: "let me know" and "I'll send another update."

It counted the 14 placeholders, flagged what was invented and what was missing, then rewrote the email in about 63 words. One honest catch: with no names in the rules, it wrote "Hello" and left the signature out. A check is only as good as the rules you give it.

Save your brief so you stop retyping it

The WHO part and your standing rules rarely change, so save them once. In ChatGPT, custom instructions "allow you to share anything you'd like ChatGPT to consider in its response", and OpenAI says they apply to all your chats. They're available on all plans: up to 1,500 characters on Free and Go, up to 5,000 on Plus, Pro, Business, Enterprise and Education (on the web: Settings, then Personalization). Other assistants have their own settings for this; check their help pages.

  • Save once: who you are, who you usually write for, your tone, your language ("Québec French, vouvoiement") and your usual format.
  • Type each time: the task and the facts.
  • Keep out: client names, account numbers, health details. Use placeholders like [CLIENT NAME] and fill them in after.

Your checklist before you press Enter

  1. Who am I, and who will read this? One line.
  2. One verb, one deliverable.
  3. The facts, pasted, with "use only these".
  4. Rules: tone, language, length.
  5. Format: the shape of the answer.
  6. The ask-first line at the end.
  7. Before it goes out: the self-check, then your own read.

Sources

  1. OpenAI Help Center, Best practices for prompt engineering with the OpenAI API (consulted Oct. 11, 2026)
  2. OpenAI Help Center, ChatGPT Custom Instructions (consulted Oct. 11, 2026)
  3. Google, Learn about prompts for Gemini in Docs, Sheets, Slides, Vids & Forms (consulted Oct. 11, 2026)
  4. Microsoft Support, Get started writing prompts in Microsoft Copilot (last updated September 2026; consulted Oct. 11, 2026)
  5. Anthropic, Claude Platform Docs: Prompting best practices (consulted Oct. 11, 2026)

Frequently asked questions

How do I use ChatGPT effectively for work?

Give it the briefing you'd give a new employee: who you are and who will read the result, one clear task, the facts to use (and only those), your rules for tone, language and length, and the format you want. Add one line asking it to ask up to three questions if something important is missing. Then read and check the result before it goes out.

What is the best way to write a ChatGPT prompt?

There's no magic prompt, but the official guides agree on the ingredients. OpenAI asks for specifics about context, outcome, length, format and style; Google lists persona, task, context and format; Microsoft lists goal, context, source and expectations. A five-part brief (who, task, facts, rules, format) covers all of them.

Why does ChatGPT give generic answers?

Because it doesn't know your situation. In our test, a one-line request for a delay email came back as a 346-word template with 14 blanks. With the facts and rules in the prompt, the same request came back as a 70-word email ready to check and send.

Does this work in Claude, Gemini or Copilot?

The same brief applies: Google, Microsoft and Anthropic recommend the same ingredients for their own assistants. The answers in this article were produced by Claude on October 11, 2026; another assistant would word them differently, so always read the result.

Can I paste client information into ChatGPT?

Only if your business has approved that tool for that kind of information. A safer habit: keep names, account numbers and health details out of the prompt, use placeholders like [CLIENT NAME], and fill them in yourself afterwards. This is general information, not legal advice.

Where should you start with AI?

The free assessment gives you, in 4 minutes, the 3 tasks to automate first in your business.

Take the free assessment

Newsletter

The essentials of AI for small businesses, once a week

The week’s articles in 5 minutes, plus a prompt and a template to use right away. Free, no ads.