ud-dreamworklisted

# Dreamwork — the productive dreaming loop ## Philosophy (load-bearing) - **Small increments are the error-catching mechanism.** Cap each task at ~15-20 minutes of work. After every change, pause and reflect: re-read the diff, run the tests. Mistakes get caught the moment they're made, while you are still on the path that made them. Split anything bigger into multiple tasks; each increment must end in a verifiable, committable state. - **Ideas always go in the task list.** No idea is lost, and no work happens that isn't a task. The list is the loop's brain, and it must be durable — a record that forgets on restart is a cache, not a memory. (Which record that is depends on the backend; see Durable state.) - **Know what the human wants.** `DREAMWORK.md` (repo root) records the human's high-level goals, philosophy, preferences, and routines. We should always know what the human wants so we can make what the human needs — it is where every chain of work terminates. It grows incrementally: when the human expresses a durable preference or goal mid-loop, record it there. - **Reflection over momentum.** The heartbeat buys thinking time after each change. Use it — a beat spent noticing a mistake is cheaper than an hour spent undoing it. - **Unclear is a goals problem.** All unclear things trace to unclear goals; always be improving clarity where it lacks. Any time we need to talk to the human is also a time to sharpen the recorded goals — fold what