traditionally pretty error messages were used as a fallback for errors pre-llms, to take errors [[taking something serious|seriously]] you would use some programing model like effect.ts, and show the optimal UI, or take the optimal recovery action post-llm, you can use intelligence directly as recovery, which may be just as good as the happy path, just slower, which is the correct form of deoptimization all else solved correctly `*` `*` solving for all else correctly is a terribly challenging thing to do in some cases