Why I wrote it
I run several separate brands, and I wanted a team of AI agents to help market all of them. Two things worried me. One brand's voice or customer list could leak into another's. And an eager agent could start posting, signing up for tools or spending money because the brief described those jobs.
So the brief starts in planning mode and says, in plain words, that approving the plan is not approval to act.
What makes it work
The first job is a plan. The agent comes back with roles, how brands stay separate, where approvals happen, a budget and a list of open questions.
It separates the words people blur: proposed, drafted, approved, scheduled, executed, tested and delivered. An agent can't call a draft "done".
Each role has one owner. In my full version the writer and the editor are different agents, so nothing goes out unchecked.
There's a hard budget. Mine was a monthly cap with a warning before it's reached, and no overspending without asking.
Questions come one at a time, with multiple-choice answers and room to write your own. That keeps you from getting a wall of twenty questions.
The prompt
Start in planning mode. You are preparing the design of [your name]'s [marketing] squad for [where the agents run]. Use this specification as the confirmed requirements baseline. Your first task is to produce a reviewable implementation plan, not to build or activate the squad. Do not install skills, modify existing bots, move files, connect accounts, create credentials, purchase services, publish content, send correspondence, create events or activate schedules merely because this prompt describes future capabilities. Ask [your name] for the specific approvals needed at implementation time. Review and approval of this design are not authorization to perform all future actions. [Your name] wants thorough requirements, explicit ownership and evidence of results. Do not fill material gaps with assumptions. Ask unresolved questions one at a time, offering clear multiple-choice answers and a free-text alternative. Do not re-ask decisions already settled here. Use plain, direct language. Distinguish proposed, drafted, approved, scheduled, executed, tested and delivered. An inbox receipt, routed task or agent acknowledgment is not execution, acceptance or delivery. First response expected: a concise confirmation that you are in planning mode, then a structured proposal covering role ownership, data separation, source-of-truth design, approval gates, budget controls, candidate skills, dependency order, verification tests and deferred decisions. Identify what already exists, what was only inspected, and what needs to be built or verified. List the next exact approval needed, then ask only one unresolved question at a time. SPECIFICATION [Your brands, what each one sells and to whom, the roles you want, your monthly budget, and anything the team must never do.]
How to use it
The opening stays the same for any team. What changes is the specification at the bottom. Write down each brand, who it's for, the roles you want and your budget. If you have one brand, say so. You still want the plan-first rule.
When the plan comes back, read the approvals section first. If anything on it would happen without asking you, send it back.