ADR Library Growth: Personal Playbook Emerges

I've been accumulating ADR files across my projects, and something interesting is happening: they're becoming a personal knowledge base that the AI understands almost as well as I do.

I really like the ADRs I've created. They help the AI understand exactly what I want, but here's the beautiful part - I'm not the one writing them. I ask the AI to generate decision records based on code I've already vetted and refined. That's leverage.

I've been running with a few ADRs about how to structure React components - specifically around RSC (React Server Components) and client components. The first attempt wasn't perfect. The AI generated components, but they didn't quite match my vision. But instead of rewriting from scratch, I tweaked the ADR descriptions with a new prompt and ran it again.

A couple weeks of that iteration, and now the AI is generating components that fit my preferences almost automatically. If I wanted to be thorough - and if I wasn't lazy - I could ask the AI to audit every existing component against the new standards and refactor everything. It would do it within the constraints of keeping everything functional. I'd just sit there hitting "continue" like a robot.

The compound effect is obvious. But I'm curious about something: what if we could take this further with finetuning or model training? If I had thousands of decisions encoded like this, could I actually train a model on my specific coding preferences?

I'd pay for that. A custom model that understands my architectural decisions, my coding style, my development philosophy. It's probably not practical at this scale, but the idea is compelling.

For now, the ADR library is working. Each project makes the next one faster because the AI carries forward what it learned from the previous ones.

That's the real goal: encoding knowledge so deeply that it becomes second nature to the system you're working with.