AI shouldn't think for the first time while you sleep

A workflow for running AI coding agents overnight without waking up to a broken codebase. The core idea: separate thinking from execution. Write structured plans in Obsidian (goal, scope, success criteria, do-not-touch, open questions), stress-test them with a 'grill-me' skill to surface vague or risky assumptions, rewrite them into boring, unambiguous execution plans, hand execution to a different agent than the one that planned, then run a single 'plan of plans' prompt that points at all prepared plans with a stop condition for anything risky. Recommends picking low-risk overnight tasks (test coverage, docs, prototypes) over high-risk ones (deploys, auth, migrations, secrets), and stresses that reviewing the diff the next morning is still the developer's job.