Blog / Build with AI /
Making AI ask you questions before it builds
You ask an assistant for “a simple booking page”. It answers in four seconds with 300 lines of code, a calendar library you have never heard of and a colour scheme that looks like a dentist’s waiting room.
It is not wrong, exactly. It guessed. When a request is vague, AI fills the gaps with the most common answer, and the most common answer is rarely yours.
The fix is cheap: tell it to ask you questions first.
Why this works
A good human freelancer does not start building when you say “a booking page”. They ask who it is for, how many bookings a day, whether people pay and what happens when someone cancels. Those questions are most of the job.
AI assistants will do the same, but only if you ask. By default they lean towards answering immediately, because that looks helpful. So you have to change the rules.
Asking questions first gives you three things:
- A better first draft, because the gaps are filled with your answers rather than guesses.
- Less rework, since misunderstandings surface before code exists.
- A clearer idea in your own head. Half the questions make you realise you had not decided something.
The basic prompt
I want to build [one sentence]. Before you write any code, ask me up to five questions that would change what you build. Ask them one at a time, and wait for my answer each time. When you have enough, summarise the plan in five bullet points and ask me to confirm.
Three details matter:
- “Questions that would change what you build” filters out filler like “what is your favourite colour?”.
- “One at a time” stops a wall of questions you will skim.
- “Summarise and confirm” gives you a checkpoint before any code appears.
If you prefer to answer in one go, swap “one at a time” for “numbered, so I can reply to all of them at once”.
What good questions look like
For the booking page, helpful questions would be:
- Who books: customers on phones, or staff entering for them?
- Do you need to stop double bookings, or is a message to you enough?
- Does it need to take payment, or just collect a request?
- Where will it live: a single HTML page, or inside a WordPress site?
- What should happen after someone submits?
Each answer removes a whole branch of the build. Question 3 alone separates a page that works in an afternoon from a project needing payment security.
If the assistant’s questions are shallow (“What features do you want?”), say so: “Ask more specific questions, based on the kind of tool this is.”
Turn the answers into a brief
Once the questions are done, ask for a short brief in your own words:
Write a half-page brief for this project: purpose, users, inputs, outputs, what is out of scope. Use my answers. I will keep it for future chats.
Save it. It becomes the first thing you paste into new chats, or into a Project if your assistant supports one. Claude Projects for long builds shows where this fits, and how to describe a tool to AI so it builds the right thing covers what a good description contains.
Variations worth keeping
The devil’s advocate:
Before building, tell me three ways this idea could go wrong or be too big for a first version. Then ask me what to cut.
The assumptions list:
List every assumption you are making about this request. I will correct the wrong ones before you start.
The user interview:
Pretend you are a first-time user of this tool. What would confuse you? Ask me what the page should say to help.
The pre-mortem:
Imagine the project has failed after a month. What are the most likely reasons? Ask me about any I can prevent now.
The assumptions list is my favourite because it works even when you are in a hurry. You can skim ten lines and spot the one that is badly off.
When to skip it
This step is not free. It costs a few minutes and a bit of patience. Skip it when:
- The task is tiny and clear, such as “change this heading to sentence case”.
- You are exploring and want a rough, fast sketch to react to.
- You already have a detailed brief.
For anything you will use more than once, or that touches a live site, it is worth the minutes. My checklist before asking AI to build anything is the long version of the same idea.
What to watch for
- Endless questions. If it asks twelve, cap it: “Ask no more than five.”
- Questions you cannot answer. Say “I don’t know; suggest a sensible default and explain it.” That is a good move for beginners.
- Forgetting the agreement. In long chats the assistant can drift away from the plan. Paste the summary again if it does.
- False confidence. A tidy summary does not mean the build will be correct. Test as usual.
A worked example
Say you manage a small shop’s website and want a page where staff can paste product names and get tidy URL slugs.
You write: “I want a page that turns product names into URL slugs. Ask me questions before you build.”
You might get: “Should accents and non-English characters be converted or removed? Do you want spaces turned into hyphens or underscores? Should the slug be shortened beyond a certain length? Should it remove words like ‘the’ and ‘and’?”
You answer in a minute. The result is a tool fitted to your store’s URL rules rather than a generic one. Without the questions, the accents rule alone could have produced a mess you only noticed after publishing.
Key takeaways
- Vague requests produce average guesses; questions fix that.
- Ask for a limited number of questions that would change the build, one at a time.
- End with a confirmed plan and save it as a brief.
- Use variations such as assumptions lists and pre-mortems.
- Skip the step for tiny, clear tasks.
Related reading
- How to describe a tool to AI so it builds the right thing
- How to ask AI to explain code line by line
What is the last thing you asked an assistant to build, and what is the one question it should have asked first?