Let the answer shape the form.
The next iteration gives the AI a richer interface vocabulary while keeping the person in control of what it understood.
the AI should figure out best line of questioning and then build a beautiful form on the go
The earlier interview work had shown why fixed questions could break a supposedly adaptive flow. I wanted to go further: the AI should choose the next useful question and build the appropriate form around it.
That is a product requirement with an engineering boundary. The model needs enough freedom to adapt, and the interface still needs predictable controls, editable information, and a recovery path.
See the original requestFrom the development record
The model proposes. The application still has a contract.
A diagram of the September local implementation direction. Earlier July screenshots would not demonstrate these new controls.
- 01
Read the answer
Preserve the information already supplied and identify the next missing detail.
- 02
Choose a valid control
Accept only supported question structures and render the appropriate input.
- 03
Keep the brief editable
Carry the answer forward without turning a model suggestion into an unquestionable fact.
This work remains in progress. The next evidence needs to show the real rendered controls and recovery paths.
The original request9 Sept 2026 · excerpt
the AI should figure out best line of questioning and then build a beautiful form on the go
Excerpted from the development conversation. Punctuation is lightly edited; the surrounding story is an edited retrospective.
01
The model needed a vocabulary of useful controls.
I was asking for more freedom in the interview, but not for an arbitrary page generated at every step. A choice group, a short answer, and a structured input each ask the person to do different work. The AI needs to select a useful way to ask, while the application retains responsibility for rendering, validation, and predictable behaviour.
That boundary is what makes the form a product rather than a stream of generated text. The same answer can influence the next question, its options, and the brief the person is building. I want those pieces to agree. A beautiful question with irrelevant choices still fails. A relevant question whose answer cannot be corrected creates a different kind of friction. The local work kept the understood details editable and exposed retry when the adaptive step could not continue.
02
Understanding must remain editable
The reported flow kept understood details editable. That matters because extraction is an interpretation. A person needs a way to correct the system without abandoning the whole request and starting again.
Failures offered retry rather than substituting stock questions. That continues the lesson from June: the interface should not disguise a failed adaptive step as a successful generic one.
03
The next problem was the pace of the conversation.
By this stage, the remaining work was not simply whether AI could return a control. City confirmation and response speed still needed refinement. A form can choose the right next question and still feel slow or uncertain while it gets there. The person needs to know whether an answer was accepted, whether the system is thinking, and whether anything is required from them.
The diagram explains the September control contract; it does not stand in for a capture of the rendered interface. The September source records local English and German browser checks and automated verification. I would next follow a fresh request from an ordinary description through correction, location confirmation, and review, watching for repeated questions and unclear waits. That keeps the work centred on the person completing the request, rather than on the novelty of a generated form.
The earlier response handling failure explains why the rendering contract matters. I return to that question in The model answered. The app threw it away.
Where this chapter landed
What changed.
The adaptive form iteration was implemented and checked locally. City confirmation and response speed remained under refinement; this entry does not claim a production release.
Added while revisiting this story · September 2026
The thinking behind the detail.
Research notes to take these decisions further. These sources were reviewed for this retrospective.
Adapt inside an explicit contract
JSON Schema documents how object properties, required fields, and extra fields are validated. A bounded control vocabulary can let an interview adapt while the application continues to validate what it knows how to render.