What building with AI actually breaks on

Everyone asks whether the AI can write the code. The answer is mostly yes. The real breakage is remembering why — and vouching for what came back.

By Arjun Venugopal · 6 October 2026 · 6 min read

Why I'm writing these. I'm a product manager, not an engineer, and I'm building a fully-local Mac app called Jiva entirely with AI. The category I'm building in is owned by companies worth a billion dollars and more. So the real question behind these posts is a simple one: can one person who can't write code build something that stands next to what those teams ship? Note that I'm not claiming it will. I'm documenting the attempt — the parts that work and the parts that don't — because that's the only honest way to find out.

Everyone asks whether the AI can write the code. In my experience that is the wrong question, because the answer is mostly yes. The thing building with an AI actually breaks on is not writing code. It is remembering why.

Here is the failure mode nobody warned me about. You have a good session with an AI, it writes something that works, you move on. Two weeks later a new session — because the model has no memory between them — cheerfully rewrites a piece you had carefully built a certain way for a reason, undoes a fix, or re-proposes an idea you already tried and rejected. The code was never the bottleneck. The bottleneck is that the reasoning behind the code evaporates the moment the session ends, and then you pay for it again and again.

Note that this is exactly why the boring markdown structure I wrote about early on exists. The decision log, the status file, the specs — they are not organisation for its own sake. They are the memory the AI does not have. When a fresh session reads them cold, it inherits not just the state of the project but the why, and stops re-litigating settled things. The primary driver of being able to build this way at all is not the model's capability, rather giving it durable memory the model itself lacks.

There is a second thing it breaks on, which is subtler: judging whether what came back is safe to ship. An agent will hand you working-looking code. As a non-engineer, I often cannot tell from reading it whether it is sound, and there is no review I would know to ask for. I have caught real bugs in my own project only by accident. That gap — between “it runs” and “I can responsibly stand behind it” — is the one I have the least good answer for, and the one I am most honest about.

So: the model writes the code. What it does not do is remember, or vouch. Those two are the actual job, and they are mine.

More soon.

Jiva is free and fully local. Record your meetings, keep speaker-labelled transcripts, and get a live assist overlay — all on your Mac, nothing sent to the cloud. Get it at jiva.works →

← All posts Previous: The markdown “brain” that let a non-coder ship →