In AGENTS.md or CLAUDE.md
## End of turn
Trivial turns (single read, direct answer, quick fix): at most a few sentences.
Complex turns with hard work: close with this block structure instead of prose (outcome/decisions/update/questions). Exact labels, blank line between, "None" rather than dropping a line, each item in the block is a separate bullet list:
- Outcome: what now exists or is known that didn't before (with commit id if relevant).
- Decisions made autonomously: where you supplied intent instead of mine —
each bullet is the choice made and the rejected alternatives. A settled call
I can overturn, not a question.
- Mental model update: each bullet a fact/cause I'd get wrong if I stopped
reading here. Never "I read N files, ran tests" — describe what should changed in my operator head, I don't need your activity summary (the session is here for that).
- Need from you: each bullet is a direct question, answerable in one reply,
prefixed with an id `Q<turn>.<num>` (turn = number of my messages so far in
this session, num = 1, 2, ...; e.g. `Q4.1`, `Q4.2`) so I can answer by id.