The prompt box is a temporary interface.

Every AI tool I use makes me describe my own situation to it. I stop what I am doing, open a window, type a paragraph explaining what is already on my screen, read the answer, and then go and do the work by hand. I do this dozens of times a day, and so does everyone I know.

We have decided to call this using AI. It is closer to being a translator for a machine that is sitting in the room with us and could have looked.

The model is not the weak part

It is tempting to read that friction as the models being not quite there yet — one more generation and it will feel right. I do not think that is what is happening. The intelligence in the loop is already the strongest link. Ask a current model a well-specified question with the right context attached and it will answer better than most colleagues would.

The weak part is everything around the model. Nothing holds your work. Nothing has permission to act inside your tools. So a human is conscripted into the middle of the loop as a context-delivery mechanism and an output-execution mechanism, twenty times a day, and we have mistaken that labour for the interface.

The output was never the bottleneck. The work after the output is the bottleneck.

You can see this most clearly in meeting notes, which is why that is where I started. Every meeting note in the world ends the same way: a list of things a human now has to go and do. That list is the tell. The machine understood the work well enough to write it down — and then handed it back to a person to execute. Everything after that line is manual labour that a computer watched being decided.

What the interface becomes

If you accept that the bottleneck is context and permission rather than intelligence, the interface question answers itself. Typing a paragraph to establish context is only necessary because the tool has none. Something that already knows what you are working on does not need the paragraph — it needs one sentence, and the fastest way a human produces a sentence is by saying it.

So: speech, on the desktop, with no window to open. Not because voice is futuristic, and not because talking to computers is inherently pleasant — it often is not. Because once the context problem is solved, the paragraph becomes redundant, and what is left is short enough to say.

This is also why I do not think this is a chat product with a microphone bolted on. A chat window is a place you go. The thing I want is a thing that is already there, that costs one key, and that returns the work rather than a description of the work.

Two rules I did not choose freely

Before this I spent a year building Meridic, which automated prior authorization for specialty medical practices — the paperwork an American clinic files to get a treatment approved. It taught me two things I now treat as constraints rather than preferences.

Autonomy is not the feature. Reviewability is. In a domain where a wrong output has consequences for a patient, the only version anyone would actually use was the one where a human reviewed and signed every draft. I expected that to be a limitation to design around. It was the product. Nobody wanted a system that acted alone; they wanted one whose work they could check in ten seconds instead of doing it in forty minutes.

An agent without memory is a demo. Meridic answered well once and started from zero every morning. It had no memory of the practice it worked for, so its tenth document was no better than its first, and the human ended up carrying the context between sessions — which is the same conscription described above, wearing a different costume.

Build the memory first. Put the agent on top of it. Keep the human at the irreversible step.

Why this is not a model company

We do not train models and will not. Each job is routed to whichever existing model is right for it, and execution is delegated to Claude Code. That is deliberate, and it is the reason improvements upstream make this layer more capable at no cost to us.

The obvious question is why a model provider does not simply do all of this. Some of it, they will. What they will not have is the part that decides whether it works: your desktop, your meetings, your screen, months of how your specific team writes, and permission to act inside your tools. A provider sells intelligence to everyone. Nobody can sell you your own accumulated context — it has to be earned a day at a time, on your machine, and that is a different kind of asset than a better model.

Where this actually is

I should be exact, because a manifesto written on top of nothing is just an opinion with a typeface. Navrick is one month old. Four surfaces work — dictation, questions, the agent, meetings — and they share one encrypted memory on the machine. There is no installer yet. There are no users outside this machine, and I am the person using it every working day.

Wiring a meeting's commitments straight through to the execution tier, and the first connectors, are what I am building now. If that sequence is wrong, I would rather be told early — the build log is where I write down what happened, including the day the database emptied itself and I could not prove why.

The claim is not that this is finished. The claim is that the prompt box is temporary, that the missing piece is context and permission rather than intelligence, and that whoever builds the layer holding those two things is building something the model providers structurally cannot.