Devblog #0

Foundry is supposed to preserve meaning

This week I kept working on Foundry.

Foundry is my attempt at understanding AI tooling by building my own version of it. Part orchestrator, part attempt at giving meaning to the changes being made.

That second part is important. Specs, memory files, context, decisions, all of that stuff exists because I do not want AI-assisted development to become random output glued together by vibes. I want the work to have some kind of thread running through it. At least that was the idea.

That is the funny part in hindsight. I was building a tool that is supposed to preserve meaning, while my own mental model was drifting out of date.

Silent internal shift

I thought my internal goal was to build stuff and understand AI tooling by rebuilding parts of it myself. Scratching an itch, more than solving a clean problem.

I watched a video today that made me realize something. I have been goal focused on my hobby projects for a while.

If I am building to learn, then the process matters. The weird detours matter. Reading the code matters. Understanding why a solution exists matters. If I am solving a problem, then I care much more about whether the problem is gone.

I wanted to reinvent some wheels. Sometimes rebuilding the wheel is how you learn what the wheel actually is.

But that only works if you are paying attention while building it. And this week made it pretty clear that I had stopped paying enough attention.

Things happened. Files changed. Specs were written. Decisions were made. The project started taking shape. And I got swept away with it.

At some point I stopped properly reading and thinking through the architecture. I got tired, skimmed through specs and decisions, and went with the flow.

That is dangerous, because the flow feels productive. It feels like the project is moving. And technically it is. But movement is not the same thing as understanding.

That was the actual trap. The project looked like it was progressing, so I did not notice that my understanding was falling behind.

The bug was that I forgot to understand

The problem was not that I used AI. The bug in my mental model was that I thought progress meant I still understood what was happening. It did not. I forgot to understand.

I forgot to understand the code, the architecture, and why one solution was picked over another. The exact choice matters less than knowing why the choice exists.

AI is very good at generating output. If you ask it to keep moving, it will keep moving. If you are tired and start skimming, it will not magically stop and force you to think.

That is how I ended up with architecture I had not properly thought through. Not because the code was magically evil, but because I let momentum make decisions for me.

We are probably building more software than ever in history, but so much of it feels worse. More code, more projects, more generated things, but less care. Less intent. Less understanding.

Slop is not just bad output. Slop is output without understanding.

That is the part I ran into this week. I was building, but I was not always thinking. I was producing, but not always deciding.

And when that happens, the goal can move without making a sound. Suddenly the question is no longer “what am I trying to understand?” It becomes “what can I get done?”

LLMs cannot think for you

I do not think the fix is to stop using AI. That would be missing the point.

The fix is thinking for yourself. Do not forget to think. Do not forget to ask why this solution and not another one.

Foundry does not need more machinery before I understand the machinery it already has. It needs me to be able to explain why something exists without hiding behind “the model made it” or “this is how other tools do it”.

I also realized that using Foundry only to build Foundry is probably a bad feedback loop this early. It sounds elegant, but it makes everything too heavy. If the tool is not ready, then trying to use it on itself turns every missing feature into a blocker. If I give up and use another tool instead, I stop feeling the pain and no longer know what to improve.

When something felt bad, I could not always tell if Foundry was bad, if the architecture was bad, if the project was too early, or if I simply did not understand what had been built.

So I probably need some kind of dummy project. I do not know what that means yet. But I know I need something outside Foundry for Foundry to push against.

Until then, the current state is less about adding features and more about getting my hands back on the steering wheel.

Slow the frak down and reverse engineer my own choices

Speed plus direction is velocity. AI can make you faster, but it cannot give you direction. If I do not know where I am going, faster only gets me lost harder.

Next I need to go through Foundry and Cerberus (Foundry’s sister project) and reverse engineer what was arguably written by me. Only after that does it make sense to decide what Foundry (and Cerberus) needs next.

Am I solving a problem, or am I scratching an itch?

Neither is wrong. But if I do not know which one I am doing, the shift can happen without me noticing.