1 Connect your agent required
One line gives Claude Code / Cursor / ChatGPT the queue tools (get_next_task, complete_task, …). Grab your token first: sidebar foot → Connect your agent → Copy token, then:
PROMPTPRIO_API_TOKEN=pp_YOUR_TOKEN bash <(curl -fsSL https://taskpeace.com/install.sh)Restart your agent, then verify: claude mcp list → promptprio ✓ Connected. Per-tool guides: Claude Code · Cursor · Cline · Goose · Warp · Continue · Codex CLI · Gemini CLI · ChatGPT · Windsurf · Zed.
2 Work the queue on autopilot
The loop — six prompts, one operating cycle. plan fills the queue with the right work → autopilot / never-stop clears it → continue resumes it if a session ever stops → capture folds any conversation back in → every session quietly trains TaskPeace itself → stop winds down clean. Paste the one you need; together they run the whole operation, with you in the loop only where your judgment is irreplaceable.
Start by filling the queue with the right work — paste the plan prompt on a project (or right after you set a goal). It turns the goal + live metrics into a ranked, well-formed backlog, routed to the right worker, so autopilot then has work genuinely worth clearing:
Or skip the goal-and-metrics route — just describe what you want and TaskPeace wraps it into the plan prompt, so your agent turns it straight into ranked, well-formed tasks. You never write the prompt:
Tell your agent to work the queue — it loops get_next_task → do it → complete_task and auto-scopes to the project your terminal is in (from the working directory). No need to say which project. Blocked tasks become a human flag, so one stuck item never stalls the run.
For the most robust run, paste our autopilot prompt — it adds hard safety gates (money / credentials / publishing stay manual), self-configures the working directory, and stops on its own when the queue is drained:
Want it to run forever until you say stop? Paste the never-stop prompt instead — same autopilot, but on an empty queue it waits (≈4 min) and re-polls rather than ending, so the moment you add a task or one comes due it's already on it. The wait keeps it cheap — it never busy-polls. Only stop / pause / halt ends it:
Session stopped — crash, context limit, closed tab, rate-limit, lost connection? Paste the continue prompt to pick up cleanly: it re-orients first (recovers any half-done task, never re-does what already shipped, makes the board true), then resumes the loop in whichever mode it was running. Resume from anywhere — nothing lost, nothing double-done:
Already in an autopilot run? You don't have to re-paste — just send q (or continue) on its own and it picks the loop back up (it leads with ▶ q · TaskPeace autopilot — resuming… so you see it caught). Only stop / pause / halt ends a run. The installer also teaches your agent the q key globally (a marked block in your CLAUDE.md), so it works from a fresh session too — every session knows it.
Done for now? Paste the stop prompt to end the session cleanly — it lands the current task (done-with-proof, or a resume note), syncs the board true, banks one learning, and leaves you a short report:
Had a valuable conversation that isn't an autopilot run — a long Claude Code session, a claude.ai chat full of decisions and ideas? Paste the capture prompt at the end and it folds everything worth keeping — follow-ups, decisions, learnings — into your queue as ranked, resumable tasks. Conversations feed the queue; the autopilot prompt works it. Nothing good is lost:
3 Run unattended / overnight optional
Keep a session alive after you close the window, and keep the Mac awake. Install once (macOS):
brew install tmux && brew install --cask ghosttyThen launch a persistent, awake session — detach with Ctrl-b d, reattach with tmux attach -t ace:
tmux new -s ace
caffeinate -dimsu claudeFor hands-off runs, let the agent act without per-step prompts — set "defaultMode": "bypassPermissions" in your agent's settings, and keep a deny-list for destructive commands. Money, credentials, new accounts & irreversible deletes should always stay manual.