Help your Agent use the full product without exposing internals.

1. Identify the active tenant and brand space.
2. Discover workflows, saved routines, destinations, provider choices, and Test & Tune capabilities.
3. Prepare a plan using the user's outcome, not an internal engine topology.
4. Explain what will run, what will be shared, expected cost, review behavior, and destination impact.
5. Request confirmation when the action spends credits, changes saved configuration, or can publish.
6. Run, monitor evidence, summarize results, and propose a focused adjustment when needed.

• Use brand, channel, topic, review, and schedule language instead of IDs and topology.
• Show a concise before-run summary and a visual or linked result after the run.
• Separate design success, review approval, scheduled handoff, and confirmed publication.
• When something fails, identify the affected workflow or destination without implying every other member failed.

An Agent cannot cross tenants, invent product facts, use unapproved assets, hide provider fallback, bypass media constraints, or turn a generation result into publication without the required authorization. It may recommend a creative choice, but product facts and execution evidence remain traceable.

When the user asks for a simple prompt or model change, prefer a focused comparison over a full production rerun. Present the changed input, variants, output differences, cost, and confirmation state. Apply the chosen result only to the scope the user approved.