08/21/2026
Fuzzgramming + Brutevibing 🤖
~~~~~~~~~~~~~~~~~~~~
This thing in building is so complex it’s nearly impossible for me to keep all the moving parts in my brain. And generally I don’t like overly complex things. But I don’t really see another way to do this and keep it flexible.
Over time I’ve simplified it a ton, but I don’t think it is simple enough. Right now I just need to use it and I’ll figure out if I can simplify it more later.
I did design it so any size organizations can use it, I want to even make it work in a single account. A remote account. A one person account. Whatever you want. For now it uses an organization by default.
But I also designed it so you can divvy out deployment privileges to different parts of your organization with different levels of permission for you know - governance.
So here’s the thing about building with AI. This is a serious app. It has to be logically accurate not vibey. I’ve built banking apps for companies with billions of dollars of assets under management. This app has some hard core logic.
Of course I wouldn’t build a banking app in bash but it was fast and lightweight for a POC. But the core logic is there to migrate the idea to some other language if you must.
So as I ponder my experience…
There’s two sides to this. On the one hand I don’t know half of what’s in it and I’m afraid to look. I will hate it and it will tempt me to immediately want to stop what I’m doing and refactor everything and that could take a lifetime given what happens when you vibe AI code.
On the other hand, I know the core logic and architecture is solid. I could be wrong and I’m sure if someone was paid a bounty they could find a flaw here and there but I think the parallel processing is pretty decent. I built it in such a way that I don’t have to fight with flock and file locks and all that.
I hope it is correct but I am still concerned about one aspect of adding errors to each node in the diagram. I think it’s right but I can’t yet say I’m certain. For some reason the project that updates the diagram data model has been the biggest problem. After getting consistent UI.
But today I got it working right end to end. It ran without the diagram locking up, fast enough, and seems to have the right data.
To the point I can USE IT.
With 5-8 terminal windows open juggling agents I can go like 100 times faster. I’m trying to quantify it in my ai-tracker repo but that’s hard to do.
The challenge is that for a single developer and especially depending on what part of the world you live in the cost is high, but for a company with a solid developer at the wheel output can go way up.
At the same time just when you think you are done, it kicks you in the teeth. It deletes key code, it creates a weird error that is so esoteric I am not even going to try to figure it out and explain it. Sometimes the models just feel haywire.
Those annoying moments are the moments I’m pondering, when you don’t know everything in the code and you don’t even want to look.
But sometimes, like today, the models are stuck and it is your best option. One little fix turned a whole bunch of nodes on my diagram green. Finding atrocious error handling and fixing it opened a flood gate of information to help solve some problems.
Can we overcome those gotchas? I don’t know. I feel like what I’m doing is fuzzgramming with a monolithic graph. I throw things at it over and over until if does down then right path and I get what I want.
Sometimes I have to slow down and literally make it write once line at a time to get it right. Brutevibing.
AI generated code is NOT pretty. It is not going to always be perfect and might generate a lot of noise but in the end you come up with something you couldn’t otherwise that “works” and produces correct outputs.
Still pondering optimizations, but today I got what I wanted. And immediately checked it into got with creds my agents are not allowed to use.
