Writing › Tag
#AI
-
What I don't let an AI agent decide
A tool for building Carpentries workshop websites with any AI coding agent. The interesting design choice wasn't the automation. It was where the agent's judgment stops and a hard rule takes over.
-
The prompt that makes a model disagree with me
A standing prompt template whose only job is to stop a second model from agreeing with my own diagnosis. Tested it on my own project-management setup and it found the weak joint in an argument I was ready to act on.
-
The single underscore that broke my task system
A CLI task manager that worked fine in my terminal silently failed inside Claude Code sessions. The cause wasn't a bug in the tool, it was a shell-snapshot mechanism that quietly drops single-underscore function names before the agent ever sees them.
-
How I run three LLMs without getting burned
Claude, ChatGPT, and NotebookLM each do different work in my writing. The thing that keeps them useful is not the model. It's a verification habit borrowed from the reference desk: every claim gets ground-truthed before it ships.
-
Three pieces I bookmarked this year that keep pointing at the same thing: a lot of the software running the world rests on one or two people, and almost no one funds them.