The markdown “brain” that let a non-coder ship
The reason I could build this isn’t the model. It’s a plain set of markdown files that give any AI session full memory of the project.
People assume the reason a non-engineer can build real software now is that the models got good. That is part of it. But the models got good for everyone, and most people who try this stall in the same place: the AI has no memory. Every new session starts from nothing. You spend the first twenty minutes re-explaining the project, and it still makes a decision you already rejected last week.
The thing that actually made Jiva possible is boring. It is a set of plain markdown files.
There is a vision doc for what the product is and why. A roadmap that lists every feature and its status as one row each. A live status file — what is built, tested, shipped, blocked, and who is on it. A numbered decision log, where every decision gets an entry with the date and, more importantly, the reason. And a wiki of short specs, one per component. That is it. No database, no tool. Files on disk that you can grep.
Note that the decision log is the load-bearing piece. Code tells you what the system does. It does not tell you why you chose this over the three things you tried and threw away. When a fresh AI session — or a different model entirely — reads these files cold, it picks up not just the state of the project but the reasoning behind it, and it stops re-litigating settled questions.
The primary value here is not organization, rather portability. Because the memory lives in files instead of a vendor's session, I can switch which AI I use mid-project at no cost. The project is not locked to any one tool.
I have since pulled this structure out into its own thing so anyone can use it. Every AI coding tool has amnesia between sessions and quietly locks your context in. This is the plain layer that fixes both. It is also the most useful thing I have learned building Jiva — more useful, honestly, than anything about the audio or the models.
More soon.