Before letting an AI agent spend metered credits on generation (video, image, or otherwise), have it first produce a written plan — scope, pieces, formats, count — and require explicit human approval before executing any of it. This turns an open-ended 'go make the content' request into a two-step gate: propose, then approve.
In practice: send a first prompt that asks for the plan while deliberately withholding the tool/connector reference (e.g., request a 7-day content plan without invoking the generation MCP), review the plan, then send a follow-up that names what's approved and explicitly invokes the connector to produce it. The checkpoint exists because agentic tool integrations often don't ask before spending — see Always-Allow Agent Permissions Risk Runaway Credit Spend for the failure mode this pattern is meant to prevent. It's distinct from Goal Mode: Closed-Loop Generate-Test-Select Content Production, which iterates after generation (generate → test → select); this pattern gates before the first generation ever happens.
Apply: when connecting an agent to a credit-metered generation tool, don't ask it to produce output in one shot. First request a plan with the generation tool unreferenced or not yet invoked, review it, then send a second prompt naming what's approved and explicitly invoking the tool to execute only that.