Personal leadership · Every phase
Asking AI a good question
You ask Claude or ChatGPT a question that gets you something useful, and you know what to check yourself.
Does this sound familiar?
- I asked for an email to a client and got something that read like a brochure.
- The answer was so general I could just as well have come up with it myself.
- The second answer wasn't any better, so I started a new chat and typed it all in again.
- There was a figure in it that sounded very sure of itself. It later turned out to be wrong.
- I pasted in a client's email and only afterwards thought: am I actually allowed to do that?
Who it's for, and when
For anyone who uses Claude or ChatGPT at work: an email, a quote, a summary, a plan. It matters in every phase. In Searching you'll mostly use it to think and to test an idea, and that's where checking counts most: a model is quick to agree with what you already think.
The model
“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”
Claude and ChatGPT know nothing about you, your company or your client unless you tell them. Whatever you leave out, they fill in with the average, and that's why a bare question comes back sounding like a brochure. Anthropic offers a test: give your question to a colleague who knows nothing about it. If they don't get what you want, the model won't either.
So a good question is the brief you'd give a new colleague: who you are and who it's for, an example of what it should look like, what you want back, and why. Then you steer it in the same chat, and you check it yourself before you use anything.
What you type goes to another company. Details of clients and colleagues (names, addresses, salaries, performance reviews, complaints) don't belong in it, unless your company has an agreement about that. On a personal Claude or ChatGPT account, what you type can be used to train the model unless you switch that off. On the business plans of both, that doesn't happen by default. For most questions you don't need names or amounts anyway: write 'client A' or 'a technician'.
Anthropic, Prompting best practices, the guide to Claude (platform.claude.com, checked on 7 November 2026). Ethan Mollick, Co-Intelligence (2024), on working with AI and staying the human who checks the work.
1. Who you are and who it's for
Two sentences about you and your company, and one about the reader. 'I own an installation company. This is for facilities managers who've known us for years.' Now the model knows which words fit.
2. An example
Paste in an email, quote or other text you wrote before and were happy with, and say what you like about it. An example shows what you mean better than a string of adjectives like 'professional but warm'.
3. What you want back, and why
Length, format and tone: 120 words at most, informal, two versions, a table. Say why. 'Short, because he'll read it on his phone' works better than just 'short', because the model then tunes everything else to that too.
4. Following up in the same chat
The first answer is a first draft. Say what's wrong and what's right: 'The second paragraph can go, the tone of version two is good.' The model builds on what's already there. Only start again if the chat has gone the wrong way, and take what you've learnt with you.
5. Checking it yourself
A model can sound very sure and still be wrong: a figure, a date, a point of law, a source that doesn't exist. Ask it whether it's sure and it'll often say yes. Check facts yourself, in your own documents or at the source. You're responsible for whatever you send.
Pitfalls
Asking whether your idea is good
Ask 'is this a good plan?' and you'll usually get a yes with a few tips. Ask where it'll fall apart instead, or have the model play the client who says no. Then you'll hear something you didn't know yet.
Everything in one question
A question with eight instructions gets you eight half answers. Do it in steps: the structure first, then the text, then cut it down. You'll also see where it goes wrong.
Watch and read
Two people from Anthropic build up a prompt step by step, from a bare first version to one with context, an example and a clear format.
Watch on YouTubeThe author of Co-Intelligence in two minutes on the skills you need to get better answers out of AI.
Watch on YouTubeThe rules from Co-Intelligence summed up in eight minutes, including that you stay the human who checks the work.
Watch on YouTubeFour tips for a better question, in Dutch and in two minutes.
Watch on YouTubeTo go further
People from Anthropic spend an hour talking about how to give a model a task, and why you should treat it as a clever colleague who knows nothing about your situation.
Watch on YouTubeRead
Co-intelligentie 2026: Slimmer werken met AI
Ethan Mollick, met een voorwoord van Alexander Klöpping · 2025 · ISBN 9789493434271
The Dutch edition (POM.press): the 2026 edition, with five new chapters and 280 pages. Mollick gives four rules for working with AI. The ones that matter most for this lesson: tell the model who it should be, and stay the human who checks the work.
View the bookCo-Intelligence: Living and Working with AI
Ethan Mollick · 2024 · ISBN 9780593716717
The English original.
View the bookApply it
- 1
Start with who you are and who it's for
Two or three sentences, before you say what you want. If you use the same sentences often, save them; in the lesson A project with your own documents you put them in once.
- 2
Paste in an example
Something you wrote before and were happy with. Take out clients' names and amounts if they don't matter.
- 3
Say what you want back
How long, in what format, in what tone, how many versions. And why: who it's for and where they'll read it.
- 4
Steer it in the same chat
Say what's good and what needs to change. After two or three rounds it's usually right. If it still isn't working by then, start again with a better first question.
- 5
Check it before you use it
Check figures, names, dates, rules and sources yourself. And read it out loud: would you say it like that?
- 6
Find out what happens to your data
If you're on a personal account, switch off training on your chats if you don't want it. In ChatGPT it's in the settings under Data controls (Improve the model for everyone), in Claude in the privacy settings (Help improve Claude). Ask what your company has agreed about client and staff data. If there's no agreement, leave them out.
Examples
The same question, wrong and right
Wrong: Write an email to a client about our price increase. Right: I own an installation company with fifteen technicians. I'm writing to our regular clients, mostly facilities managers of office buildings, to say our hourly rate goes up from €68 to €72 on 1 January. They've known us for years and appreciate that we're direct. Write an email of 120 words at most, informal, with no apologies and no marketing language. Give the reason in one sentence: wages under our collective labour agreement (cao) are going up. Keep it short, because most of them read it on their phone. Below is an email I sent last year and was happy with. Keep to that tone. [paste the email] Give me two versions. Then, in the same chat: Version two is good. Make the first sentence shorter and take out 'we understand that'.
Following up instead of starting again
This is too long. Cut it in half and keep the last paragraph. The tone is too formal. That's not how I talk to a client I've known for ten years. What am I missing? What would the client ask next? Now write it from the client's side: what do they read first? Give it as a list of five points instead of running text.
Checking: five questions
Do the figures come from my own documents, or did the model make them up? Are the names, dates and amounts right? Is there a law, rule or source in it? Look it up yourself. Does the text promise something we can't deliver? Would I say this myself?
What doesn't belong in it, and what to write instead
A client's name and address: 'a client, an office in the area'. A client's complaint email with their details: sum up the complaint in your own words. A colleague's salary or performance review: 'a technician who's been with us for two years'. Medical or personal details: leave them out, even without a name. A contract with a confidentiality clause: leave it out, unless your company has a business plan with an agreement covering it.
Do this in your own business
Everything up to here is free. What you fill in for your own business, you keep in Boldcaster.
- You look back on a situation from your own work. Only you can read what you write.
- You set a date with yourself, and on that day the app asks how it went.
- You assign the lesson to your team and see who has done it.
- The quiz counts towards your progress, and you see which lesson comes next.
Four weeks free. You'll land straight in this lesson.
Quiz
6 situations. You need 5 of 6 right. You can try again at any time.
1.You ask ChatGPT for a job advert for a planner. You get text that would fit any company. What do you do first?
2.The second answer to your question is better, but the tone still isn't right. What fits this lesson?
3.In a note for your leadership team, Claude writes that the legal notice period for a permanent contract in the Netherlands is two months. It sounds sure. What do you do?
4.A client sends an angry email with his name, address and a story about his wife being ill. You want help with the reply, and you're using your personal ChatGPT. Your company has no agreement about this. What do you do?
5.You want to know whether your plan for a new service holds up. Which question gets you the most?
6.You want the newsletter to sound the way you always write. What works best?
Nothing is saved here. In the app, the quiz counts towards your progress.
Read next
- A project with your own documents · You set up a project in Claude or ChatGPT that knows who you are and how you work, so you don't have to explain it every time.
Not sure where to start? Take the growth scan