been thinking about documentation wrong this whole time. i treated it as a description of the code. when an agent does the work, it should be a panel where i watch the thing think.

code is the only output i get. the process evaporates. which libraries it rejected and why, where it hit an incompatibility, what it learned reading someone else’s repo, which parts are still half done. next session it doesn’t remember any of that either, so it re-researches what it already knew last tuesday.

so: research/ for what was considered and on what criteria, blockers/ for what is broken right now, learnings/ for what came out of it. cross-linked, newest surfaced. current state, in files, written by the thing doing the work.

i’m not fully sure what this turns into yet. right now it feels like a workshop wall with everything pinned up where you can see it.