How Work Moves

How work moves

1. Read the current task, authority, instructions and relevant source. Check the revision and destination.
2. Find existing work before creating a new packet. Claim shared work using the team's established mechanism.
3. Prepare a bounded plan and acceptance test. Set the total budget and stop rule before dispatching workers.
4. Do the work in the system that owns it. Save a checkpoint before a session ends or context is lost.
5. Give a reviewer the artifact, request, source and standard. Preserve disagreements and repair real defects.
6. Deliver only through the authorized route. Read the result back at the destination and record what was observed.
7. Update the same task record. Save the lesson and next action; close only when the required checks are complete.

A handoff is accepted when the next owner can open the same canonical artifact and revision, understands the current state and acceptance test, and has the authority and tools to continue. Include queued work, remaining budget, blockers and disagreements. Until acceptance, the current coordinator retains responsibility.

A judgment brief should contain the exact ask, decision needed, permitted actions, constraints, evidence links, alternatives, uncertainties and required output. It saves attention without replacing the evidence. A reviewer may request the source passage that could change the answer.

One durable brief, one result

Separate apps do not automatically share all conversation context. Save a brief the next worker can actually open. The brief points to the original evidence; it does not replace or filter out contradictory evidence.

Task ID and revision: [same ID for retries]
Outcome owner / receiving worker / reviewer: [names or functions]
Requested result and acceptance test: [observable result]
Inputs: [authorized source links, revisions and retrieval dates]
Allowed actions, audience, budget and stop rule: [bounds]
Evidence, alternatives, uncertainty and dissent: [links]
Last verified checkpoint, in-flight work and next action: [state]

Receiving worker: open this record and the needed source. Confirm
the revision, permissions and next step. If unreadable, record BLOCKED.

Return to this same record: recommendation and reasons; exact changed
artifacts; checks and failures; remaining contradictions; usage or
UNKNOWN; next action. The owner accepts or reopens the result.

Sending a handoff is not acceptance. Until the receiving owner confirms the record and can continue within its authority, the current owner retains the outcome. Stop the affected dependent action when evidence contradicts a claim or a required check fails. Preserve the failure and escalate to the owner or reviewer; only a reserved human decision goes to the human. Check the destination before retrying an action that may already have happened.

One morning snapshot

A useful coordinator gives the owner one view of the same task records, rather than several competing briefings. Start with the top three highest-value actions that require the owner, each with evidence, consequence, deadline and a recommended decision. Keep agent-owned production below them.

Then list agent-owned work with its next checkpoint, and real blockers with the exact missing decision or capability, evidence, recovery owner and next check. A missing reply is not permission, and a service error is not automatically a login problem.

Current adoption status: Muse reports that a morning snapshot is configured and a routine card has been visible. Independent settings readback and the first unattended delivery have not been verified; response errors and the latest update's acceptance remain unresolved with the setup owner. This is a requested operating format, not a claim that delivery works. Reuse an existing coordinator routine; a guide edit does not create a schedule.

Can the next worker open the actual artifact?

Before dispatch, the sender names the exact artifact and revision, where the receiver will read it, what may change and where the result must return. Both workers first make a harmless read in the intended destination with their own account. Check the file itself, rather than only its linking page.

1. Open the current task record and the actual source or file.
2. Name the revision and check the source that controls the next action.
3. Confirm permitted audience, action, tools and budget.
4. Check for an existing claim, in-flight action or result before retrying.
5. Accept the bounded next step, or record the exact access or authority blocker.

A Library item, cloud file or local path that opens for the sender may be outside the receiver's scope. Keep the task blocked at that boundary. Use an existing approved route or obtain the specific missing permission from its owner. Keep the same job and record; a dispatch alone is not an accepted handoff.

Return the result to that record with artifact revision, checks, unresolved issues and next action. The receiving owner or reviewer opens the returned artifact where it landed. Only then may the outcome owner accept it.

Routine Claude status belongs on the existing shared GitHub job or handoff, where Travel Dot and Happy review it. Dennis's inbox is for messages that need his judgment, not project management. Escalate a bounded consequential decision, missing authority or material contradiction after the agents attempt the supported route. Use the original client channel for authorized client communication.

A written routing rule does not prove a live scheduled prompt loaded it. Check the actual account-specific prompt, its saved revision and the next real run. Current runtime adoption of this routing remains unverified until those tests pass.

A small worked loop

Example only: turn an approved workshop recording into a useful teaching page. Travel Dot frames the lesson and acceptance criteria. Happy finds the existing task and assigns an available worker. The worker reads the authorized recording, drafts from its evidence and flags missing facts. A separate reviewer checks against the recording and publishing standard. The existing publish desk receives accepted files, publishes within recorded authority and returns the live URL. The owner checks the page and records what was learned. The same task ID connects the loop.

Make a real audit reproducible

Use an already authorized client task and its actual evidence. Pin the source artifact, date, allowed scope, rubric and report revision. Another worker should be able to repeat the checks and explain any changed result. Missing evidence stays unknown; a draft report is not client delivery.

Record who opened the source, who produced the report, who checked it, which exact artifact the next worker opened and what was accepted. Show public readers only client-approved facts and permitted artifacts. Until the receiver and owner complete this loop, label the demonstration pending; a sample template or saved report does not prove the handoff.