LLM behavior improvement

multiple captures around making LLMs better at their jobs:

codebase scanning for weak spots: noticed Claude and Codex failing on Prisma — scan the codebase for things the LLM is not confident about, then note exactly how to deal with them in AGENTS.md. "kinda like spec but for weird parts of codebase."
acting without enough info: LLMs often behave poorly because they act without enough information. could be solved with better prompting/telling users to give more info, or with a dataset training models to guess intent intelligently and refuse unspecced tasks.
graph thinking for LLMs: current approaches crudely approximate human thinking (x then y then z) but miss backtracking and organization. proxy: linear text with organization indicators. noted counterpoint: attention mechanism already integrates everything non-linearly, "so possibly moot."

connects to the AGENTS.md research and the overnight app grinder which depends on LLMs being reliable.

timeline