How I use AI
What I hand to models, what I keep, and the rule that shows up in everything I build.
Rules belong in code, not prompts
A prompt that says "don't do X" is a request. A check that throws the output away when it does X is a rule. In call_human the tutor is told never to write your code, but the thing that actually stops it is a server-side scan that discards any reply containing code you didn't write. Orderable reports an unlisted allergen as unknown, never as safe. OpenPango blocks a customer email that cites an ETA the carrier never gave. Models are good at the work. They shouldn't be the last line of defense.
What I hand off
- Scaffolding, boilerplate, and the fifth version of a form I've already built four times.
- First drafts of tests, docs, and migrations, which I then read line by line.
- Searching unfamiliar codebases and docs to find where to start.
What I keep
- Learning fundamentals. If I can't write it without help, I don't understand it yet.
- Decisions about what to build, what to cut, and what to stop.
- Anything I'd have to defend in an interview, an exam, or a postmortem.
Tools
Claude Code and the Claude API for most building, OpenAI-compatible APIs and Groq where speed matters, and local models through Ollama and LM Studio when I want to see what runs on my own hardware.