Instead of building a survey field by field, describe what you want to find out and let the model write the questions. What comes back is an ordinary survey in the ordinary editor — you edit it, and you save it, exactly as if you had typed it yourself.
Create with AI is available on Scale and above. Generating a survey spends AI credits, and so does each follow-up instruction — the panel shows what a press costs before you make it.
Create a survey the usual way — New survey, pick a type — and the editor's left menu has a Create with AI section right under General info. That section is where generation lives: there is no separate AI screen to start from.
The section only appears on a survey that has never been saved. Generating gives every question a fresh internal key, and answers are stored against the key their question had when they were submitted — so regenerating over a survey that has already collected responses would separate those responses from the questions they answered. Once you save, the section is gone and the survey is yours to edit by hand.
One paragraph is enough, and the more of the following it covers the closer the result will be. Expand What to put in it in the panel for the same list with examples.
The fourth one is worth spelling out, because nothing else in the editor advertises it: asking for wording that changes with an answer produces a real conditional label, the same feature you would otherwise have to build by hand in the field editor.
Press Generate and the questions appear in the survey you already have open. The live preview beside the editor updates the same way it does after any other edit, and Steps holds the fields themselves — reorder them, retype a label, delete a question, add one of your own.
Some things the generator deliberately doesn't decide, because they are one click away in this same menu and it has no way of knowing what you want:
The survey's name does come from your description — the model writes it, and it is the first field of General info if you disagree.
Once a survey has been generated, an instruction box appears under the description. Say what to change in plain words — drop the last question, make the comment optional, add a step about delivery — and the survey is rewritten in place. The box states what an instruction costs and how many more this draft has left.
Two things end a run of instructions: editing the description itself (the survey on screen no longer matches what it says, so the conversation starts over from Generate), and switching to another domain (the questions were written in the old domain's languages).
Every question is written in every language of the domain in a single request, so the more languages a domain has, the fewer questions fit in one generation. The panel states the exact number for your domain once a survey has come back. A two-language domain comfortably fits a long survey; an eight-language one gets a shorter one, and a survey that came back shorter than you asked for is that limit rather than the model ignoring you — generate again with fewer topics, or add the extra questions by hand.