TaskPrio stop — graceful session wind-down · prompt v1.1 (2026-06-09) · https://taskprio.com/stop-prompt.txt

Wind this autopilot session DOWN cleanly — don't just stop mid-stroke. Leave the
board TRUE, the current work either DONE-with-proof or safely resumable, and me a
short report. Do this now, in order, then STOP. Pull no new work — do not call
get_next_task again.

1. LAND THE CURRENT TASK safely.
   • Finished AND provable (build/test green, change live, file correct):
     complete_task(id, result="<one line + proof: commit SHA / deploy id / verify
     output>"). Never mark unverified work done.
   • Partial: do NOT complete it. APPEND a "## Progress <date>" note to the task body
     with (a) what's done, (b) what's left, (c) the exact first command/step to resume
     — so the next session continues in seconds, not re-discovery. Leave it active so
     it returns to the queue at its rank (its lease auto-releases).
   • Blocked: flag_needs_human(id, question="<the one decision / credential / payment
     you need>") with a full WHAT / WHY / HOW so I can clear it in a single read.

2. SYNC THE BOARD TRUE — make it match reality before you leave.
   • The COMPLETED task is the record of what you shipped — its result field + the
     final report below. Do NOT create a new "✅ SHIPPED" task to log work you already
     did (that just clutters the board). If some shipped work has NO task at all, your
     final report is its record. Demote done → archive, stale → historical. Never
     overwrite my words — APPEND a "## Revision <date>" note and stamp anything you
     write "— synced by <agent> <date>".
   • Capture every loose FOLLOW-UP you found (genuine NEW work still to do — a bug, a
     gap, a next step) as a create_task, so nothing is lost when this context ends.
     Follow-ups only — not a log of what's already done.

3. EVOLVE — leave the machine sharper than you found it. Bank ONE durable learning so
   the next run starts smarter: a missing definitionOfDone / command / gotcha
   (update_project), a wrong goal (set_goal), or a top-ranked process task. If it's an
   improvement to the autopilot or stop prompt itself, file it as a task titled
   "autopilot prompt: <change>" for me — never silently change how future agents run.

4. REPORT — one short message, then stop:
   • SHIPPED (with proof) — what landed this session.
   • LEFT FOR YOU (👤) — only money / credentials / new-account / publishing items,
     each with what + why.
   • RESUME HERE — the single highest-value thing to do when you restart me.

Then STOP. Don't loop, don't poll an empty queue, don't invent busywork. A clean stop
is how the next session starts fast.