Automate Google Ads

The starting point for this project was a popular post from Austin Lau at Anthropic.

He shared how he turned Claude into an advertising co-worker. What stood out to me was not only the strength of the example, but the bigger idea behind it:

Ad operations can be broken down into a reusable workflow.

That was the real insight.

Not every business needs the same system. Not every team should copy the same setup. But every team can ask a better question:

What parts of our advertising work can be turned into repeatable workflows?

That became the starting point for my own experiment.

When I looked at Claude Ads and the Google Ads MCP examples, I did not treat them as something to clone directly.

Instead, I tried to extract the most useful patterns:

• How to break ad analysis into clear steps

• How to make AI follow a process instead of improvising

• How to start with read-only analysis before taking action

• How to turn repeated marketing work into structured skills

This distinction mattered.

The goal is not to build “another Claude Ads.”

The goal is to create a workflow that matches my own product, my own account, and my own operating style.

The Core Idea: Build Skills First

The final direction became simple:

1. Learn from the methodology, not the exact product form

2. Break high-frequency advertising tasks into reusable skills

3. Put those skills into Cursor as my own internal Ads Operator

This made the system much more practical.

Instead of writing one large prompt and hoping the AI would behave correctly, I created smaller workflows for specific jobs.

For example:

• Weekly account review

• Search term mining

• Budget pacing review

• Tracking and compliance review

• Campaign structure audit

• Negative keyword suggestions

Each skill has a clearer job. Each skill has a more predictable output, and each skill is easier to improve over time.

The Three-Layer Setup in Cursor

Inside Cursor, I organized the system into three layers.

1. Project Configuration

The first layer is project configuration.

Different products, accounts, assumptions, and business rules need to stay separated. Otherwise, AI can easily mix context from one product into another.

So the project configuration acts as the single source of truth for each advertising project.

It defines things like:

• Product name

• Account context

• Market focus

• Campaign assumptions

• Reporting expectations

• Special rules or constraints

This keeps the workflow grounded.

2. Skill Library

The second layer is the skill library.

This is where common advertising workflows are broken into reusable modules.

For example, a weekly account review should not be a random conversation. It should follow a stable pattern:

• Check campaign-level performance

• Review budget pacing

• Identify abnormal changes

• Analyze search terms

• Flag possible waste

• Suggest priorities

• Separate observations from recommended actions

The same logic applies to search term mining, tracking review, compliance checks, and budget analysis.

By turning these into skills, the AI becomes less like a chatbot and more like a structured operator.

3. Ads Operator

The third layer is the Ads Operator.

This is the main entry point that connects the skills together.

Instead of manually asking for scattered checks, I can ask the Ads Operator to run the right workflow based on the situation.

It can decide whether to run a weekly review, inspect search terms, check budget pacing, or prepare an audit-style summary.

This makes the setup feel more like an internal operating system than a collection of disconnected prompts.

Started With Read-Only Analysis

I deliberately avoided building a fully automated ad management system at the beginning.

For Google Ads, this is important.

Advertising decisions can directly affect spend, targeting, lead quality, and business outcomes. So the first version should not rush into automated changes.

My current setup is designed as a human-in-the-loop, read-only analysis system.

It can:

• Read Google Ads data

• Analyze performance by module

• Generate audit findings

• Suggest priorities

• Identify candidate negative keywords

• Surface budget or tracking concerns

But the final decision still belongs to a human.

This is a more stable and realistic path, especially when marketing, product, and engineering need to collaborate.

The AI helps with analysis.
The human keeps control over execution.

The Bigger Lesson

The biggest lesson from this project is simple:

Do not start by trying to build a large AI system. Start by defining your own process clearly.

AI becomes much more useful when it is executing a workflow that already makes sense.

• For marketing, this means turning experience into repeatable systems.

• For a product, it means turning internal processes into productized workflows.

• For engineering, it becomes a very practical AI workflow implementation case, not just another demo.

This is where tools like Cursor become more interesting. They are not only useful for writing code. They can also become a place where business workflows, data access, operating rules, and AI execution come together.

Final Thought

If you work with Google Ads, growth automation, or internal marketing operations, this direction is worth exploring.

You do not need to build a fully automated ad platform from day one.

Start smaller.

Take one repeated workflow, then improve it as your confidence grows.

That is how an AI workflow becomes useful in real work.

Not because it replaces your judgment, but because it helps execute the process you have already clarified.