9 June 2026

Although Bliptown proudly states it has no AI, I still employ AI to help me write it. I always keep it at a distance, letting it show me how something should work and then I will rewrite it in my own style. I never let agents run on my local machine (and never on the production server).

This poses a problem, because for AI to be helpful it needs to read the existing code. One way this works is for the AI to operate on the same platform as the code is hosted, like GitHub. I had got a free GitHub Copilot student subscription that initially gave me access to the latest ChatGPT and Claude, but then Claude was taken away, and then all but the free versions of ChatGPT were gone too. Also I felt uncomfortable about using GitHub.

I wanted some way for AI to read my code, but I didn’t want to use GitHub any more, and I didn’t want to allow it to operate on my computer.

Most AI CLI tools work on a permission system, but the permissions are built into the tools, so it seems trivial for the agent to just change or ignore them. I looked at various options to sandbox the agent, like the (deprecated) macOS sandbox-exec CLI, but it was too complicated and wouldn’t help on my Alpine Linux machine.

Then I found llm, a Python tool that takes the prompt from STDIN and, sends it to a model, then prints the response to STDOUT. Beautifully minimal. No AI runs locally. No complicated sandboxing system needed.

I wrote a PROMPT.sh that would print high-level model instructions about Bliptown, then the name and content of each file. I can pipe this into llm then provide an additional prompt on the command line and voila — the model has full knowledge of the entire project.

Currently listening: Grouper – Ruins