Directing an AI marketing agent: direction, skills and memory
Directing an AI marketing agent comes down to three tools: a written direction that says where to optimize and what to protect, reusable skills that carry your methods from task to task, and memory that keeps what earlier runs learned. Keep approvals on while the agent learns your account.
1. A written direction
Most disappointing agent work is not wrong, it is aimed at the wrong thing. A direction fixes that in a few lines: where to look for improvement, what must not change, what to avoid trying.
- Where to optimize — "focus on non-brand search and Performance Max".
- What to protect — "never pause brand campaigns; keep the retargeting budget".
- What to avoid — "no broad-match expansion this quarter".
- Write constraints as rules, not wishes. "Never pause brand campaigns" is checkable; "be careful with brand" is not.
2. Skills: write the method once
A skill is a reusable method — how your team audits search terms, how it structures a weekly client report, which checks come before a budget increase. Written once, it can be pinned wherever it applies instead of being re-explained in every goal.
- 1Start from a task that went well and write down the steps that made it good.
- 2Save it as a skill — or let the agent draft one from a rough idea, then edit it.
- 3Pin it on the projects and tasks where it applies. In Agent Planners a project can pin up to ten skills, and a project's direction can itself be saved as a skill for reuse.
- 4Revise the skill when the method changes, rather than adding exceptions to every goal.
3. Shared context and memory
An agent that forgets last week repeats last week's mistakes. Memory is what earlier runs learned about an account — its targets, its quirks, the experiment that failed — recalled when the next run plans.
- Account facts beat general advice. "This account's conversions arrive two days late" changes how every later report should be read.
- Review what was learned. Remove a stale fact before it misleads a plan.
- Keep direction and memory separate. Direction is what you want; memory is what is true. Mixing them makes both harder to correct.
4. Keep approvals on while it learns
A new direction or skill is a hypothesis about how the agent should work. The approval cards are where you see whether it holds: each proposed change, its reasoning and its numbers. By default every change waits for your approval; turn on auto-approve for ordinary changes only once the proposals consistently match what you would have done — and even then, deletes, live sends and code still wait for a person.
Frequently asked questions
- How do I tell an AI marketing agent what to prioritize?
- Give it a short written direction — where to optimize, what to protect, what to avoid — phrased as rules it can check, and attach reusable skills for the methods you want followed.
- What is a skill for an AI agent?
- A reusable written method — an audit procedure, a report structure, a pre-flight checklist — that is pinned on the tasks or projects where it applies, so the agent follows it without being told again.
- When should I stop approving every change an agent makes?
- When its proposals consistently match what you would have done. Until then keep the default approvals; changes a person must decide, such as deletes and live sends, wait for a person even with auto-approve on.