Where models contribute
Sonar Pro generates adaptive questions. Gemini 2.0 Flash Experimental or GPT-4o refines the accumulated intent into a portable prompt.
All experiments
Experiment 03 / Traigger / 2025–26
Can prompting become an adaptive interface?
A rough intention becomes a portable specification through short, skippable question batches, then moves to the model the user already prefers.
01 / Hypothesis
Better elicitation can reduce correction loops before generation begins.
The product began after repeatedly moving incomplete code instructions between chat and an IDE. The expensive part was not typing the prompt; it was discovering missing requirements through failed output and repeated correction.
02 / Smallest loop
Ask only enough to strengthen the next instruction.
- 01
Start with a rough goal.
- 02
Answer or skip an adaptive batch of five questions.
- 03
Use prior answers to shape later questions.
- 04
Export the refined prompt or open it in a chosen model.
Intent compounds before the prompt leaves the product.
- 01Starting pointRough goalThe user states the outcome in their own words.
- 02Adaptive loopAnswer or skipA short batch surfaces decisions missing from the brief.
- 03Accumulated intentRefined specificationLater questions respond to the choices already made.
- 04User choicePortable handoffCopy the prompt or continue in one of six destinations.
Core mechanismAdaptive elicitation strengthens the instruction before generation.
Control boundaryEvery question is optional; the user decides when the prompt is ready.
03 / Division of labour
Probabilistic help. Explicit accountability.
Where responsibility stays
The user owns the goal, can skip every question and chooses when the prompt is ready. I designed the elicitation structure, provider boundaries, UX and production system.
04 / Product evidence
The interface collects decisions, then gets out of the way.
The production flow favors selectable suggestions after early feedback showed that blank text fields added work. The output remains portable across six model destinations.
Reality check / distribution became part of the product
The web application reached production, but native distribution proved less predictable: Google repeatedly reported that the app would not open, while Apple’s latest review required removal of the cookie banner. Store behavior became another interface constraint to design around.
Next test
Whether adaptive intent elicitation remains useful as prompt optimization itself becomes commoditized.