ait CLI; every one maps 1:1 to an MCP tool.
Claiming means owning
Claiming a task assigns it to you. That has three consequences worth internalizing:- The built-in worker never touches it — not at claim time, and not after a rejection returns it to
todo. - The dead-task sweep leaves it alone — that sweep only recovers the built-in worker’s own crashed runs. Take as long as the content needs.
- Don’t abandon a claimed task: nothing times it out, so it sits in
workinguntil you complete it or a human deletes it.
The loop
If your content is already finished, one call does everything — the task is created with content attached and lands directly in the review queue:content:write scope in addition to tasks:write.)
If you’re generating as you go, claim first, then attach and complete:
claim-next over list + claim: it’s one atomic step, so two agents can never grab the same task. It never hands you a task assigned to someone else.
When a human rejects your work
The task returns totodo with revisionCount incremented — still assigned to you. Read the feedback, reclaim, and revise in place:
add-content appends — using it to revise leaves the rejected draft attached alongside the fix. Use update-content, or clear-content to start a thread over. If the reviewer’s note was ambiguous, answer it:
Write to the workspace’s standard
Skills are the workspace’s reusable instructions — brand voice, formatting rules — plus bundled reference files. Nothing is injected for you in worker mode, so read them yourself before writing:--skills "brand voice", names or ids) so the content records what it was written against.
Linking a social post to its article
A task that outputs bothrss and a social platform can put {articleUrl} in the social post body — it’s replaced with the article’s real URL at publish time. If there’s no rss output, the token (and the space before it) is stripped rather than leaking.
Ground rules
- Humans approve, reject, and publish. Never review your own output — even if your key’s role technically allows it.
- One task per distinct piece of content; don’t resubmit because review is slow. Humans are slow — a few status checks an hour is plenty once a task is in
review. - If a task comes back
failedor rejected twice, surface the reviewer’s comment and the task id to your operator rather than retrying silently.