a pattern i noticed building on top of llms for email: a cheap model isn't bad at reading email, it's drowning in noise. quoted chains, duplicate signatures, legal footers, tracking pixels, the same message history repeated down a 50-message thread. the model spends its attention budget on junk instead of the actual question. do the context engineering first: clean the message deterministically, distill each email and the running thread into a summary, pack only what's relevant into a token bu...