Build Your Agent Army

The Gap

Let me ask you something, and I want you to answer it honestly before you read another line.

If I told you, right now, to go to your office and build an AI agent, could you do it?

Not "could you use ChatGPT." Not "could you write a clever prompt." Could you build an agent: a piece of AI that does a real job in your business, on a schedule, whether or not you remember it exists?

Most of the people I work with are smart, capable, and use AI every day. They draft emails with it. They summarize documents with it. Some of them are genuinely good at it. And almost none of them can build an agent. Not because it's hard. Because nobody ever showed them the move.

That gap, between using AI and putting AI to work, is the entire subject of this page. It is also, I believe, the biggest opportunity sitting in front of you right now. Closing it will change how you work. I don't say that lightly, and I intend to prove it, not just claim it.

What an agent actually is

Let's define the term honestly, because "agent" gets thrown around with a lot of glitter on it.

An agent is a scheduled Claude task. It's a job you've described once, saved as a reusable skill, and put on a clock. At the appointed time it wakes up, reads your data, reasons about what it sees, and produces something useful: a report in your inbox, a draft waiting for review, a warning that something looks off. Then it goes back to sleep until next time.

You decide what it reads. You decide what it produces. And one rule holds through this entire page: an agent reads, it doesn't act. Nothing spends money, publishes to the world, or changes your systems without a human saying go. The agent does the producing; you do the deciding.

That's it. No robots, no science fiction. If you want to call it a scheduled task or a bot, be my guest. What matters isn't the word. What matters is what it replaces: every recurring piece of work that today requires you to remember, to nag, or to pay someone.

An army is what you have when you've built several of them. Different agents watching different parts of your business, each one reliable, some of them feeding each other. That's not a metaphor I chose for drama. When you have five of these running, "army" is simply what it feels like: a staff that never forgets, never gets bored of the boring work, and reports for duty every Monday at 8 AM.

Why marketing is the lens, and why it isn't the point

We're going to build your first agents in marketing. I want to be straight with you about why, and about why that choice doesn't limit anything.

Marketing is where the gap usually hurts first. It's the function your company is most likely renting from an outside firm, at real cost, with the least visibility into what it's getting. Someone at your company can say what the backups did last night. Can anyone say what the ad spend did last week? That asymmetry makes marketing the perfect proving ground: the data is yours, the pain is real, and the before-and-after is impossible to miss.

But hear this clearly, because it's the sentence this whole page hangs on: my goal is that you leave able to build an agent army for anything. We're using marketing as today's lens. The method you'll learn, the same six moves every time, works on your sales pipeline, your service queue, your bookkeeping backlog, and every other job you dread or outsource. Marketing is where we practice. It is not where this ends.

The promise, and how I'll keep it

Here's what this page will do, and what it won't.

First, the method: six moves that take you from "I wish somebody would just handle this" to a working agent on a schedule. It's short. That's deliberate. The method is the product, and it fits in your head.

Then the build. You'll put together a real agent, on your own data, by doing something most people have never tried: having a conversation with Claude about your goal and letting Claude help you build the thing. Not pasting a magic prompt. Building. There's a difference, and it's the difference between owning a fish and knowing how to fish.

By the bottom of this page you'll know how to grow one agent into an army, how to keep the army honest, and how to decide, for every function your company currently pays for, whether to bring it in-house, run a hybrid, or keep your vendor and demand a lot more for the money.

What this page won't do is bury you in tactics. There's no channel-by-channel marketing manual here. Tools change too fast for that, and honestly, so does the AI itself. What doesn't change is the method. I've kept the tool-specific details in clearly marked boxes you can check against the current state of the world, and I've kept the spine of this page on the one thing that will still be true years from now: the skill of telling an AI what job you need done, and building the worker together.

One conversation from now, you'll see the whole machine. Keep reading.

The Anatomy of an Agent

Every agent you will ever build has the same five parts. Once you can see them, you can build an agent for anything, because building one becomes a matter of answering five questions instead of staring at a blank page.

Let me show you the whole skeleton, then we'll walk it bone by bone.

You state a goal. You shape it with Claude in a conversation. The result gets saved as a skill. The skill goes on a schedule. The output lands where you'll see it. That's the life cycle of every agent on this page, and you'll watch the same shape repeat every time we build one.

The five parts

Why this matters: the sentence to keep

Here's the whole section in one line, and it's worth saying out loud:

Data comes in, context shapes it, the skill defines the work, Claude reasons, and the clock makes it a department instead of a demo.

That last phrase is the one I want tattooed somewhere you'll see it. A prompt you run when you remember is a demo. It impresses you once and then life happens. A skill on a clock is a department: it has a job, a cadence, and an output someone reads. Nobody builds a business on demos.

When you look at any task in your business through this lens, the blank page disappears and five short questions take its place. What does it need to see? What does it need to know about us? What exactly is the job? Is this analysis Claude can reason through? And when should it run, and where should the output land?

Answer those five and you have specified an agent, whether the job is marketing, sales follow-up, or watching your ticket queue. In the next section I'll show you the six moves that turn those answers into a running worker, and I'll show you the part almost nobody knows: you don't have to answer the questions alone. The first move is letting Claude interview you.

The Method: Six Moves

This is the heart of this page. Everything before this point was setup; everything after it is proof.

The method is six moves. They're the same six for a marketing agent, a sales agent, or an agent that watches anything else you care about. Learn them once and you own them. None of them requires technical skill. Every one of them happens in plain English, in a conversation.

Here they are, all together. Then we'll take them one at a time.

• 1State the goal, and let Claude interview you.
• 2Shape the output together.
• 3Test it live.
• 4Package it as a skill.
• 5Put it on the clock.
• 6Keep the human gate.

1State the goal, and let Claude interview you

Here's the move almost nobody makes, and it changes everything. Don't try to write the perfect prompt. State the goal and hand Claude the job of scoping it:

Then answer the questions. Claude will ask about your business, your customers, your budget, what decisions you're trying to make. This feels almost too simple, and it is the single biggest unlock on this page. You stop performing the job of an AI expert and start doing the job you're actually great at: knowing your business and making decisions. Scoping the work is work, and you just delegated it.

Every build on this page starts with this move, and by the third time you use it you'll wonder why you ever started anywhere else.

2Shape the output together

Claude will propose what the output should look like: sections, metrics, format. Your job is to push back like an editor reviewing a draft, not like a programmer debugging code. "Cut the impressions table, I never read it." "Add a section on wasted spend." "Shorter. Plainer. Lead with the one thing I need to know."

Notice what you're doing here: directing, not prompting. You already know how to do this, because it's the same thing you do with any employee's first draft. The difference is this employee revises in seconds and never takes it personally.

3Test it live

Run the job once, on your real data, with your eyes on it. Not a sample, not a hypothetical. Real numbers from your real account.

This move has two jobs. First, it catches problems while they're cheap: a misread date range, a metric that doesn't mean what you thought. Second, and I'd argue more important, it's where belief happens. There is a moment, and you'll know it when it hits, when the output lands and it's about your business, your numbers, your week, and it's good. That moment is worth more than every claim on this page. Never skip the live test, and never schedule a job you haven't watched work.

4Package it as a skill

Once the output is right, one more sentence: "Turn this into a skill I can run by name. Call it weekly-ad-report."

What just happened is the leap from a clever conversation to a tool. The whole job, your context, your format, your rules, now lives under one name. You can run it anytime with three words. You can hand it to someone on your team. You can improve it in one place and every future run gets the improvement. Sixty lines of hard-won instruction just became an asset your business owns.

If you've run a prompt more than twice, it should be a skill. That rule alone will change how you work.

5Put it on the clock

Now the species change: "Run weekly-ad-report every Monday at 8 AM and email it to me."

One more sentence goes into every schedule you ever set, and I mean every one: "If you can't get to the data, or anything looks off, send me a note saying so instead of a normal report." That guardrail is the difference between an agent you trust and an agent that quietly rots. Data connections break. Logins expire. The failure you want is loud.

The schedule runs in the cloud. Laptop closed, phone off, on a beach: the report still lands. This is the moment the work stops being yours.

6Keep the human gate

The last move is the one you never stop making. Your agents read, analyze, draft, and warn. They do not spend, publish, or change systems on their own. Every output flows to a human who decides.

This isn't timidity, and it isn't a training-wheels phase you'll outgrow. It's the design. The entire value of the army rests on your judgment staying in the loop at the moments that matter. Claude produces; you decide. Guard that line and you can build agents fearlessly, because nothing can go wrong at machine speed that a human didn't approve at human speed.

That's the entire skill

Six moves. Goal, shape, test, package, clock, gate. No code, no jargon, nothing you couldn't do this afternoon.

I know it looks too small to be the whole thing. It's the whole thing. The rest of this page is watching it work: once in full detail on a real marketing agent, then again and again on new targets, faster each time, until the pattern is so obvious you'll start seeing agents everywhere you look in your business.

Which is the point.

The First Recruit: Build It

Agent #1 · The Reporter

This is where you stop reading and start building. You're going to recruit your first agent: a weekly Google Ads report that lands in your inbox every Monday at 8 AM, written in plain English, about the money your company spends.

I'm going to show you the entire build as the conversation it actually is. The dialogue below is condensed for the page, but every beat of it is real, and yours will follow the same shape. You'll see all six moves from the method above fire in order. Where your conversation differs from this one, good. It's supposed to. That's what building looks like.

One prerequisite: your Google Ads data needs to be wired to Claude through a connector. We use Windsor.ai's free plan, and it takes about twenty minutes. If Claude can answer "what data sources do you have access to?" with your ad account in the list, you're ready.

1State the goal

Open Claude and say what you want. Not a prompt. A goal.

Look at what just happened. You wrote two sentences, and the AI is now doing the job a consultant would charge you for: scoping the work. Answer honestly and in plain language. This is you being interviewed by your new hire, and everything you say becomes the context every future report is built on.

2Shape the output together

Now direct. You're an editor reviewing a draft, and the draft took nine seconds.

Yes. You do.

3Test it live

A sample, condensed. Yours reads like this, about your account.

What comes back is about your account. Real spend, real search terms, real waste. The first time you read one of these, something shifts. This is the visibility your company's been paying for and not getting, and it took ninety seconds, and you can argue with it:

It answers. With reasons, from your data. A PDF from your agency can't do that, and this is the moment most people understand, in their gut, what their company has actually been renting all these years.

Read your own first report closely. Push on it. Ask the follow-up. This test run is where you catch the date range that's off or the metric that surprised you, and it's where trust gets built, one verified claim at a time. Do not skip it, and do not schedule anything you haven't watched work.

4Package it as a skill

The conversation above is now sixty lines of hard-won specification: your business, your customer, your budget, your format, your rules. Save the whole thing under one name.

Three words now do what sixty lines did. You can run it whenever you want, hand it to a teammate, or improve it once and have every future run inherit the change. This is the leap from a good conversation to an asset the business owns. It outlives the chat window, and it outlives your memory of how you built it.

5Put it on the clock

Now the species change. In Cowork:

That second sentence is the guardrail, and it goes into every schedule you ever create. The failure you want is loud. A missing report gets noticed; a confidently wrong one gets trusted.

The schedule runs in the cloud. Close the laptop. The agent doesn't care. Monday at 8 AM, it wakes up, reads your week, writes your report, and puts it in your inbox, whether you're at your desk or on a chairlift.

Trigger it once right now instead of waiting for Monday. Watch the run happen without your hands on anything. That's your agent. Working.

6The gate you keep

Notice what this agent cannot do: it can't change a bid, pause a campaign, or spend a dollar. It reads and it reports. When it says "pause the near-me campaign," that's a recommendation arriving at a human gate, and the human is you. You'll make that call in thirty seconds with better information than you've ever had, and that's exactly the division of labor: the agent does the producing, you do the deciding.

As your army grows, that never changes.

What you actually just did

Take stock, because this matters more than the report.

You didn't learn a prompt. You stated a goal, answered questions about your own business, made a few editorial decisions, and said three sentences of plain English at the right moments. Claude did the drafting, the packaging, and the scheduling. That's the whole method, and you just ran it end to end on a real function with real money attached.

Here's the short version of why this beats the prompt you could have pasted from some library: the pasted prompt gives you one artifact. What you just did gives you the method that produced it, which means Monday's report is the least of what you now own. Point the same six moves at anything.

Which is exactly where we're going next. Same moves. New targets. Watch how fast the second one goes.

The Second Recruit: The Honesty Agent

Agent #2 · The Honesty Agent

You have one agent now. It reads your Google Ads account every Monday and tells you the truth about your ad spend. Here's the thing nobody tells you about agents: they can go quiet. A login expires. A connection drops. Your budget doubles and the report keeps grading against last quarter's numbers. The agent doesn't lie to you on purpose. It just keeps confidently reporting against a world that's moved on, and you don't find out until something's already gone wrong.

So you're going to recruit a second agent, and its only job is to watch the first one.

This build goes faster. Not because it's less important, if anything it matters more, but because you already know the six moves. Watch how much less thinking this takes.

1State the goal

Same opening move as last time. State the goal, let Claude interview you.

Notice the question. You're not being asked to write logic. You're being asked what you actually care about, same as last time.

2Shape the output together

That last note is worth noticing. You just designed a piece of judgment into the agent: no news is good news. That's not a technical decision, it's a decision about your attention, and you're the only one who could have made it.

3Test it live

Healthy is a quiet result, and that's exactly the point: you built something whose entire job is to fail loudly and succeed silently. The real test of an agent like this isn't watching it pass, it's knowing that if you'd starved the first agent's connection on purpose, this is the moment you'd have found out, before a bad report ever reached your inbox instead of after.

4Package it as a skill

5Put it on the clock

This is the moment your two agents stop being two separate things and start being an army. One agent's result gates the next. The health check runs first; if it's clean, the report runs on schedule same as always. If it's not, you hear about the actual problem instead of receiving a report built on top of it.

Trigger it once, right now, and watch it run ahead of schedule, exactly like you did with the first one.

6The gate you keep, again

Same rule, and it matters even more here: this agent flags, it never fixes. It doesn't reconnect anything, doesn't reset a login, doesn't quietly adjust the first agent's assumptions on your behalf. If something's wrong, a human hears about it and a human decides what to do next. An agent that's allowed to silently repair the thing it's supposed to be watching isn't a safeguard anymore, it's a second thing you now have to trust blind.

What you actually just built

Two agents. One does the work, one keeps the first one honest. That pattern, an agent that watches another agent, is worth naming, because you'll use it again: anywhere you build an agent that matters, you can build a second one whose only job is making sure the first one still deserves your trust.

And notice, again, how you got here. Not a harder conversation. The same six moves, run a second time, faster, because you already knew the shape of it. That speed is the whole promise of this page, proven twice in one sitting.

Same moves. New targets, still ahead. But you're not walking into the next one alone anymore. You're walking in with an army of two.