Your AI Needs a Manager, Not Better Prompts.

After studying how Anthropic engineers use Claude, I identified their biggest differentiator. It's a process I call a Bottom-Up AI System, and the concept is simple.

Right now Claude is reactionary. You ask something, it reacts, then shares the answer.

We want it to be proactive. It shares the answer with you BEFORE you ask, and then you review the output.

Here's why this is so powerful and how you can steal the same strategy.

Anthropic's Engineers Already Do This

You don't have to take my word that this works. Just look at how the people building these models actually use them.

Cat Wu is a product lead at Anthropic. Here’s what she said about what Claude does for their team:

Their system looks at the work being done and suggests the automations to make it better. That's not a tool waiting for you to pick it up. That's your employee noticing what needs to be done and reporting it directly to you.

But there’s another part that’s just as important, and that’s setting up WHEN these suggestions and reports come to you. Boris Cherny is one of the creators of Claude Code, and he explained the morning routine that starts his day:

Anthropic’s team has the same tools you have. The only difference is that they built a proactive structure.

And until you build a similar structure, you’re going to be stuck in the Grinder Overload Trap.

The Grinder Overload Trap

If you’re reading this, you’re likely a continuous learner who’s willing to work hard to get things done. A term that people call “grinders,” something I’ve personally hung my hat on during my professional career.

But the reality is, this personality trait can be the biggest reason you fail.

For example, let’s say you have 5 AI tools you use. You’re the connection to all of them, and your work ethic fills in the inefficiencies of your system.

Ask yourself: if you went on vacation, would ANYTHING in your AI system still run without you?

If not, you’re working on a Top-Down System. You manage it all. What you need is a Bottom-Up System, which is empowered by having an AI Manager report directly to you. Below is a visual of how you’re likely working now versus how you will be working once you implement these systems.

The Top-Down System is not a tools problem. It's a structural problem. And you can't prompt or grind your way out of it.

The move is to stop doing everything by hand. Let the cracks surface so you can see them. Then build the Bottom-Up structure that can keep everything moving WITHOUT YOU.  

That’s what Anthropic does, but the truth is, so have I. But before we get to that, you have to enhance you system…

Which brings us to today's newsletter sponsor, Bolt.new, and their new product Bolt Slides. This is the new way to create ANY Slide Deck you’re working on. It’s as simple as saying, "Build me a 5-page deck about an AI automation I can build for a client" in their web app or directly in Claude with their integration. Bolt builds a deck in seconds.

Now it does help you move faster, which is important, but that's not why I like it. I like it because it makes decks BETTER in 3 ways:

1. You can embed anything. Live data, a clickable diagram, the world’s your oyster.

2. It looks native everywhere. You can view it on any device (Phone, tablet, etc.)

3. You visualize as you create. Bolt lets you see concepts early so you can iterate quickly without wasting time.

[Click here to start using Bolt Slides today.]

The Manager Playbook

I spent years as a COO scaling a tech company worth well over $25M. And while working with and learning from the same operators who scaled Facebook, Amazon, and Netflix, I learned something.

Great management is boring and repetitive.

It's clearly defined roles and a steady reporting rhythm. And as the manager, the reports land with me when I expect them to, whether I chase them or not.

Managing your AI is the same job. Here are the two levers to set in place:

1. Definition. What exactly needs to happen.

2. Cadence. When it lands, set to repeat.

It’s exactly what Cat Wu and Boris Cherny say Anthropic does. And it takes time. You don't hire someone and get perfect work on day one. You coach them. You correct the first few outputs. Same with a Claude Skill you create to automate a task.

Now you may be thinking, “This makes sense conceptually, but how do you actually set it up?”

Here's what it looks like in practice, using a system I actually run.

The Four Layers

To create a proactive AI system that reports to you, there are 4 layers you need to build. I’ll walk through handling customer service requests.

Workers. Individual Claude skills that each do one job well. One classifies requests into buckets, /classify-requests . One drafts replies in my voice /respond-to-customer. Each has a SPECIFIC job to do, and that’s that, clearly avoiding scope creep.

The key here, like training an employee, is that you HAVE to monitor these skills’ outputs and provide feedback directly. After each iteration, try this prompt

Orchestration skill. One skill that runs the workers together, like a team lead. It fires the classifier, hands the output to the reply writer, and packages the result.

The process of having one skill use many worker skills is a process I call “Skill Chaining.”

Routine. The cadence. A scheduled trigger that runs the orchestration skill at 7am, every day, without me touching it.

Within Claude, there’s a feature called Routines where you can set up and easily reference the Orchestration skill you created.

Manager loop. If each orchestration is a team lead, the manager loop is the department manager that reports the results to you. The best way to do this is to have all of your routines write to a specific file that the manager then reads at a specific cadence. For me, I love using Slack as the location where the “Manager” reports to me in Slack.

Here’s a graphic to help you visualize it better:

Notice the arrows point up. The work flows to you. You're not digging through six tools to check if anything happened.

The bad news: this is not something you can instantly build and get right. It requires coaching, same as onboarding a real hire. But once it runs, it RUNS. And you're finally at the top.

Build Your First System This Week

To get it started for yourself, pick a single thing you do by hand every day. Let’s say it’s replying to emails.

First, create a skill. A skill is just a folder with a short file that tells Claude how to do one job. Start with a single worker. Paste this in:

Swap the emails and buckets for whatever you handle by hand every day: new leads, support tickets, job applications, whatever.

Second, give it a cadence by setting up a Claude routine.

That alone is a proactive system.

When you’re ready to include more workers, add one skill that runs your workers together. Then add a manager report skill that reads the results, flags anything off, and sends a five-line summary to you in Slack. That summary will become the one thing you open.

That's what creating a proactive Bottom-Up AI System looks like. Not a better prompt, but a structure that reports to you.

Define the job, set the cadence, and let the work come to you. That's your first hire.

If you got this far, quick life update from me!

Yesterday - I turned 30. Today - this Newsletter hit ~10,000 subscribers & we got our first brand partnership. And in 2 days - I will compete in the Lake Placid Ironman.

29 for me was crazy, so I figured it was only right to start 30 with a 2.4-mile swim, 112-mile bike and a 26.2-mile run.

Appreciate all of you who read and support my content 🙏 

Let’s keep building. I’ll see you in the next one

Whenever you're ready, here’s how I can help:

For Business Owners & Executives:
1. Apply for my Executive AI Coaching Program: Linked here
2. Want to build a SaaS product without hiring a CTO? Linked here 

For Everyone:
3. Use BuildPartner.ai to build faster with Claude Code (try free): Linked here
4. Forwarded this email? Sign up here