​
Completed state

At the end of this guide, a launch-announcement Issue has one assignee, an inspectable Agent Run and output, a structured reviewer decision, and an accepted done state. The Issue itself contains enough evidence for a later reader to understand what happened.

​
Prerequisites

Prepare a bounded outcome, acceptance criteria, one eligible assignee, and a reviewer when independent judgment matters. Issue status meanings and allowed transitions are defined in Issue statuses; this guide does not create another status definition. Every Issue keeps one assignee. Assignment to an Agent wakes it when the Issue is actionable rather than in backlog. Agent execution checks out the Issue with Run ownership, and a named reviewer cannot be bypassed by assignee completion.

​
Run the launch announcement

  1. Create Draft the launch announcement with the audience, required setup command, claim boundaries, output format, and reviewer acceptance criteria.
  2. Leave it in backlog while any of those inputs still need a decision. Move it to todo only when another owner can start without asking what the task means.
  3. Assign the writing Agent and name the marketing reviewer. Assignment wakes the Agent; do not start a duplicate Run manually.
  4. Let the Agent check out the Issue and produce the Markdown file. Inspect the Agent Run, output link, validation, cost when available, and remaining risk.
  5. Move the Issue to in_review only when the output is ready for judgment. Reviewer routing gives the reviewer the next action without reassigning the implementation work.
  6. The reviewer records one structured decision with a comment. Acceptance can complete the Issue. Requested changes return the next step to the assignee; a blocked decision keeps the missing input visible.
A normal comment preserves context but does not wake the assignee. When a specific Agent must act on a comment, use an explicit wake mention. The mention does not reassign the Issue.

​
Success signal

The done Issue shows the original request, one current owner for each stage, the Agent Run, Markdown output, validation evidence, review comment, structured decision, and any follow-up work. No one needs a separate terminal transcript or private message to reconstruct the result.

​
Recovery

If the Run fails, keep the Issue open and attach the failure evidence. Use blocked when the next action depends on access, a person, a dependency, or an external condition, and name who can unblock it. If the assignee finishes without a sufficient close-out signal, Rudder may schedule bounded same-Agent follow-up rather than hide the gap as success. If review output exists but no structured decision was recorded, keep the Issue in review and route it back to the reviewer. A closed Issue can be reopened only through an explicit reopen comment, which leaves durable evidence.