How I use AI

If I can do something myself, I do it myself so I reach for LLMs when it can help me decompose large projects and ideas and/or get to a useful starting point faster.

I do not treat its output as finished work or as a substitute for understanding the result.

Programming

Generated code is always reviewed, and I often rewrite it. This is because models have some very specific, and to me, often frustrating patterns. I make this fixes to keep context lean and focused on what the agent needs to solve the problem. Code can also become so complex and unmaintainable with LLMs. I’m guilty of it on my own as well so thorough review is important for maintainability for me.

Tooling

Codex with Sol is my main orchestration harness. I use it to inspect repositories, plan changes, implement some features, run checks, and review the result.

Pi, usually with 5.6 Luna & Terra, handles one-off, well-defined tasks.

I’ll talk to ChatGPT (chat, not work) when a conversation is a better fit than a repository session.

Other

Herdr is a multiplexer to help coordinates multiple coding agents when a problem can be split into independent pieces.

Thndrs is my own agent harness and a place to experiment with how those workflows should behave.

Writing

I use AI more sparingly for prose than for code. It does more work to help organize notes, some proof-reading or help me understand ideas. I do not publish generated prose as-is.

Anything AI-assisted is edited and extensively rewritten so the judgment, wording, and final claims are mine. Plus, LLM assisted writing is just bad.

Research and learning

I use AI-assisted web search extensively as a jumping-off point. In my opinion this is its greatest utility.

I follow those leads to primary sources and I do not rely on a generated summary as evidence.

This website

AI tools assist with the site’s code, review, and early copy drafts under the same rules above. Like all my projects, I review every change before it is published or deployed.