The rule
Agents may create anything and publish nothing. Every piece of AI-generated content lands in a REVIEW state inside your scheduling tool, never directly on a client's feed. The approval click is the product — it's what clients are paying you to exercise judgment on.
When to run it
Any time a client calendar drops below two weeks of runway, or when onboarding a new client and you need thirty days of on-brand material fast.
The play
1. Feed the agent the brand facts: voice, pillars, offers, no-go topics, platform list. 2. Have it research what's currently working in the niche — hooks, formats, trending audio — and write platform-native drafts, not one caption pasted five times. 3. Render or attach media per post (templates beat one-offs; a branded template set pays rent forever). 4. Stage everything as REVIEW items on the calendar with suggested slots. 5. A human sweeps the queue in one sitting: approve, edit, or kill. Ten minutes for thirty posts. 6. Track what got killed and why — feed that back into the brand facts so the kill-rate drops every month.
Where it breaks
Letting "just this once" auto-publish creep in — one bad AI post to a client account costs more trust than a hundred good ones earn. And skipping the feedback loop: if the same edits keep recurring, the brief is wrong, not the model.