Interfaces that think with you
Most AI products race to give you an answer. The more interesting problem is building one that asks you a better question.
There is a default shape that AI products keep collapsing into. A text box, a send button, and a model that responds as fast as it can. It is a good shape — it made the technology legible to millions of people almost overnight. But it encodes an assumption worth questioning: that what a person wants is an answer, delivered quickly, and that the interface's job is to get out of the way.
Sometimes that is exactly right. If you need a function rewritten or a paragraph translated, speed is the whole product. But a surprising amount of what people actually bring to these tools is not a request for information. It is something closer to a half-formed thought they are hoping to finish. And for that, an instant answer is not just unhelpful — it is the wrong move entirely. It ends the process at the moment the process was supposed to begin.
The question is the product
I built Inkling around this inversion. It is a journaling app with an AI companion, and the core design constraint is that the companion asks one question at a time. Not a summary of your feelings. Not five bullet points of advice. One question, and then it waits.
That constraint sounds like a limitation and behaves like a feature. When a system responds to a tangled thought with a confident synthesis, it takes the thinking away from you — and the synthesis is usually wrong in ways you cannot see, because it is built from three sentences you typed while unsure of what you meant. When it responds with a question, it hands the thinking back. The article you eventually produce is yours, because you wrote your way into it.
An answer closes a loop. A question opens one. Most of the value in reflective software lives inside the loop that stays open.
This is not a claim that models should be less capable. It is a claim about where capability should be spent. The hard engineering in Inkling is not in generating text — that part is nearly free now. It is in deciding what to ask, when to stay quiet, and when the conversation has enough substance to become something finished.
Translation, not automation
The same pattern shows up in work that looks unrelated on the surface. Gēnova TCM deals with Traditional Chinese Medicine — tongue diagnosis, herbal formulas, five-element wellness. Auspice works with BaZi, the Chinese system of reading a life from the time of a birth. Both are dense traditions with centuries of internal structure, and both are usually encountered either through a practitioner or not at all.
The temptation with systems like these is to automate the expert away — feed in inputs, return a verdict. That produces something that is simultaneously overconfident and shallow. What these traditions actually need from software is translation: a way in for someone who has no vocabulary for it yet, that does not flatten the thing itself in the process. The interface has to carry the humility that a good practitioner would carry.
So the design question stops being "how do we compute the answer" and becomes "what does this person need to understand before an answer would even mean anything." Those are very different products, and only one of them is worth building.
Constraint as a feature
ImmeCam pushes the idea to its most literal form: one cinematic frame. Not a burst, not a roll — a single deliberate image. Every photographer knows the feeling of returning from somewhere with four hundred photographs and no picture. Removing the option to spray is not a missing feature; it is the entire proposition.
What connects all four products is a belief that the interface is not packaging around a model. It is where the intent actually lives. A model can generate a thousand words about your week. Whether those words are worth anything depends entirely on what the interface asked you first, and how much room it left you to answer badly, change your mind, and try again.
What this asks of the builder
Designing this way is slower and less demo-friendly. A product that waits is hard to show in thirty seconds. A product that asks one question looks, in a screenshot, like a product that does less. The payoff is not in the first session — it is in whether someone comes back, and whether what they made with it feels like theirs.
That is the bet behind everything in the lit chamber: that human intent and machine intelligence produce the most interesting results when the machine is built to make room for the human, rather than to finish their sentences.