My Policy on AI

All the writing on this blog is mine.

I write this blog because I enjoy writing and I’d like people to read what I’ve written¹. I’ll use LLMs to grammar and spell check, although I find they rarely do a better job than just re-reading what I’ve written.

I think I’m a reasonably good writer. I’m not against LLMs writing for me per se, but as of now they write badly and with lots of cliches, and I am against bad writing.

The illustrations on this blog are AI-generated

I think I’m a reasonably lousy visual artist. I don’t make any money off this blog. If I didn’t image-gen some illustrations I wouldn’t pay anyone to draw them for me, so there just wouldn’t be any. A good image or visual pun can punch up an article: I think the Shelf of Despair in this post is pretty funny, and I prompted Nano Banana for a while to get it how I liked it.

I use LLMs as tools to help me write code

I’ll use them for what they’re good for. That’s changing weekly, but at the time of this writing in early 2026:

• They’re excellent at generating boilerplate with my codebase or doc for reference.
• They’re pretty good at mechanical refactoring
• They can code up a good starting point for unit test cases
• They’re perfect for getting started on a greenfield app in common languages and frameworks

In general, if the thing I’m trying to do would involve me googling doc or Stack Overflow answers and 90% of the result would be copy/paste from what I looked up, an LLM can do it significantly faster than I could.

I treat a coding agent like an inexperienced but sharp junior dev, but one that doesn’t learn day-to-day. You can give it a well-defined task and it’ll give you a pretty good result, but if you ask too much of it without looking over its shoulder you’ll end up with a real mess.

I generally review all its generated code as if I was reviewing a PR, and I’ll rewrite it or pick-and-choose from its results as I see fit.

When I use AI for anything non-trivial I will credit it.

Simple, easy, honest.