Write prompts that work

A prompt is simply the message you type in the chat. You don't need special words or technical vocabulary: the agent is built to talk with non-developers, asks questions when something is unclear and proposes options as buttons. But the clearer your message, the fewer back-and-forths you need — and the fewer credits you spend. Here is what works.

How the agent reads your first message

For anything bigger than a one-line fix, the agent does not start coding straight away. Its first reply:

1. restates what it understood in a sentence or two;
2. asks up to three questions if something important is missing (brand, colours, pages, language…), often with clickable options;
3. proposes a plan of 3 to 6 steps, each one something you can see in the preview.

Then it builds one step per reply and waits for you. So your first prompt does not need to contain everything — but every fact you give up front is a question it won't have to ask. See Approve the plan, step by step.

The five ingredients of a good first prompt

1. What it is. « A booking site », « an online shop », « an internal dashboard ».
2. Who it's for and what they do there. « Clients book a haircut and pay nothing online. »
3. The pages or features. List them: home, services with prices, booking form, contact.
4. Your real content. Business name, city, opening hours, prices, the tone you want. Real details beat placeholders.
5. The look. A style in plain words (« minimal and bright », « dark and premium »), your colours, or a site you like.

A booking site for my barbershop "Fade Factory" in Lyon. Pages: home with a big photo and our motto "Sharp cuts, no wait", services with prices (cut 25 €, beard 15 €, cut + beard 35 €), a booking form (name, phone, service, day and time slot), and contact with our address 12 rue Mercière and opening hours Tue–Sat 9:00–19:00. Style: black and white with a mustard accent, bold titles. Site in French.

Tip

write the site's language explicitly (« site in French », « site in English »). The agent answers you in your interface language, but the site's own text follows what you ask for.

Give it references

• A site you like: paste its address and say what you like about it (« the layout of the home page », « the colours »). The agent can study a reference site's fonts, colours and sections and take inspiration from it.
• Your own files: on Starter and Pro you can attach PDFs, images and documents to a message, or add them once in Files (« Project files ») so the agent sees them on every prompt — a menu, a price list, a logo, a brochure.
• Photos: share your own, or let the agent find free stock photos that match your topic.

Make the home page feel like https://example.com — big photo at the top, short sentences, lots of white space — but keep our colours and our own text.

Asking for changes

After the first build, most of your prompts will be changes. The rule: say what you see and what you want, in the user's words, not how to code it.

• Point at the place. « On the services page, the second card… » is faster than « the card ».
• Describe the result. « Make the booking button bigger and green » rather than « change the CSS ».
• One topic per message. Three unrelated changes in one prompt are three chances for a misunderstanding. Group small edits on the same page, separate the rest.
• Say what's wrong, precisely. « When I click Book on mobile nothing happens » is something the agent can investigate; « it doesn't work » is not.

On the mobile view, the menu covers the page title. Can you move the title below the menu and make the menu button a bit smaller?

Answering the agent's questions

When the agent asks something, it usually shows the options as buttons: click one, or type your own answer in the box below (« Pick an option above, or type your own answer… »). If none of the options fits, just say so in your own words.

What to avoid

• « Build everything at once. » A huge request is split into steps anyway, and a single reply that tries to do it all hits the per-message limits. See How credits work.
• « Rebuild / restart / refresh to fix it. » The platform runs and reloads everything; describe the symptom instead.
• Secrets in the chat. Never paste passwords, API keys or tokens. Put them in Variables — see Environment variables.
• Repeating the same failing request. If something didn't work twice, change the angle: describe the problem differently, go back to a previous version, or ask the agent to propose options.

FAQ

Related

• Start with your idea
• Make it yours
• Versions: undo anything
• When something breaks