miniblog.

After further playing with my LLM project, I'm surprised how hard it is to tune with system prompts. My agent kept saying "obviously" even though my prompt said "helpful, professional". Eventually I found that a "courteous" prompt gets "of course" which is better but not ideal.

I've realised that AGENTS.md / CLAUDE.md solve two distinct problems: * It tells the LLM the conventions of the project, avoiding follow-up prompts. * It precomputes a summary of the project, saving a bunch of searches at the start of a session.

Amazing project that trained the GPT-2 language model on text adventure games, producing a game that allows arbitrary prompts! Game: https://aidungeon.io/ About: https://pcc.cs.byu.edu/2019/11/21/ai-dungeon-2-creating-infinitely-generated-text-adventures-with-deep-learning-language-models/ (Ironic that a text game has large GPU requirements to run too.)