Trust the parser, not the prompt: what running a cheap LLM in production taught me

I run an LLM in the core loop of a small production app: - daily writing practice that remembers which mistakes you keep repeating. The memory side of it is written up in error memory under the hood.