QUICK STARTS
Start with a bounded task loop, put the same shape into a reusable plan, then try a scheduled maintenance loop with weekly review. Each page gives you copy-paste commands, one prompt, and the log from a run that actually happened.
Each of these was verified by actually running it — the linked log is from that run, starting with the starter task.
QS1 · TASK LOOP · 10 MINUTES
Start with one red command, give the agent a three-pass budget, and come back only when the check is green, blocked, or exhausted.
Outcome: A tiny starter task changed from `status=todo` to `status=done`; the check went from red to green and the diff was recorded.
QS2 · CONTRACT + PASS LOG · 30 MINUTES
Move the task, check, budget, blocked rule, and log path out of the prompt and into files a reviewer can audit without replaying the transcript.
Outcome: The same starter task ran under a filled task contract; its pass log records the state read, check evidence, written state, and terminal verdict.
QS4 · RECURRING MAINTENANCE · ONE WEEK
Use a conservative link audit as a recurring one-pass loop: every firing writes a report, appends a full pass-log entry, and escalates source-replacement decisions instead of guessing.
Outcome: Two real report-only firings were run from this repo. Both appended full-format pass entries; the second included a synthetic dead-link probe to exercise the escalation path without public or destructive changes.