Paperclip: persistent work, not another chat persona
Updated October 30, 2026. Educational guidance; use current official documentation and your installed version.
Paperclip is the optional coordination layer in Adam's setup. OpenClaw remains the conversational gateway; Paperclip tracks specialist assignments and runs; Mission Control presents private operational state.
What belongs in a task?
Use a persistent task when work needs an owner, source material, an acceptance test, review, or follow-up beyond the current conversation. Give it a bounded outcome, not an open-ended mission to “improve everything.”
Outcome: a private, review-ready answer to this course question.
Sources: the current thread, relevant assignment, and existing Sent/Drafts.
Owner: teaching specialist.
Limits: one thread; no sending, extensions, or grade changes.
Acceptance: verified facts, saved draft or explicit blocker, and artifact path.
Budget: one finite run; no recursive delegation or automatic paid fallback.
Who does what?
The current role guide distinguishes portfolio owners, shared specialists, and the human reviewer. Sixteen configured roles do not mean sixteen agents work continuously. Automatic chat-to-task routing must be deliberately implemented; matching a name in a prompt does not establish a handoff.
Install for your environment
Use the official Paperclip repository and installation instructions. Verify the current package, runtime, authentication, and supported adapters. Keep board credentials server-side. Do not copy Adam's private settings, internal addresses, or authentication files.
A dependable handoff
- Detect a relevant change with deterministic code where possible.
- Check for an active or already completed assignment with the same source revision.
- Create or trigger one bounded task with an accountable owner.
- Verify a real execution starts; a todo item alone is not progress.
- Inspect the artifact and compare with the source system.
- Reconcile its status with any work completed directly through OpenClaw.
- Notify the human only when there is a result, decision, or actionable blocker.
Use daily caps, one active run per specialist, skip-missed scheduling, and coalescing where supported. Avoid installing competing timers in multiple systems. See proactive workflows.