How Anthropic's Growth Marketer Built a Google Ads Plugin With Claude Cowork

If you're running paid search campaigns, you know the drill: log into Google Ads, download search terms, filter for negatives, check bids, adjust budgets. Repeat weekly. It's tedious work that eats hours.

Austin Lau, growth marketer at Anthropic, got tired of it. So he built something better — a custom Claude Cowork plugin that does the heavy lifting for him.

The Problem With Manual Campaign Management

Managing Google Ads at scale means repetitive tasks:

• Mining search terms for negative keywords
• Checking impression share by day
• Optimizing budgets across campaigns
• Writing weekly performance reports

Individually, each task takes 15–30 minutes. But as campaigns grow, you're looking at hours every week spent in the interface — clicking through menus, exporting CSVs, manually filtering data.

"I realized I was doing the same analysis every week, just with different numbers," Austin said in his viral X thread. "The logic never changed. Only the data did."

The Solution: Skills as Workflows

Instead of repeating himself, Austin encoded his workflows into Claude Cowork Skills — structured prompts that tell Claude exactly how to approach each task.

His plugin includes several skills:

• Mine Search Terms — queries campaigns via GAQL, filters for low-intent searches, categorizes negatives by reason (wrong intent, competitor brand, research-only, too generic)
• Budget Optimization — analyzes spend patterns, suggests reallocations based on performance
• Weekly Review — pulls key metrics, generates a summary with trend analysis
• Impression Share Check — gets daily impression share data (something the Google Ads mobile app doesn't support)

Each skill contains specific steps, evaluation criteria, and output formats — basically giving Claude the same instructions he'd give a junior marketer.

What the Output Looks Like

When Austin runs the search-term-mining skill, Claude:

1. Queries campaigns through the Google Ads MCP connector
2. Filters search terms based on the skill's logic
3. Evaluates each term against his methodology (spend, conversions, search intent)
4. Generates a summary with top spending negatives and reasoning
5. Exports a CSV ready for Google Ads Editor bulk upload

The key difference? Claude follows his exact rules every time. No shortcuts, no assumptions — just the methodology he designed.

Mutations Require Approval

Here's the important part: for anything that changes data (adding negatives, adjusting bids, modifying budgets), Claude asks first. Always.

Austin calls this "approval-gated mutations" — the AI proposes, humans approve. It's the pattern every performance marketer should use when delegating to AI.

The Mobile Superpower

Here's what makes this really powerful: Austin uses Dispatch (Claude's mobile app) for quick operations.

From his phone, he can:

• Pull impression share by day
• Get budget recommendations
• Apply changes without opening his laptop

The Google Ads mobile app can't do impression share by day. But with his custom skill + MCP, Claude can — right from Dispatch.

How to Build Your Own

You don't need to be a developer. Here's the pattern:

1. Connect Google Ads via MCP — use a community server like Adspirer or itallstartedwithaidea's google-ads-mcp for write access, or Google's official server for reads

2. Write your first skill — document your methodology in prompts. What steps do you follow? What criteria matter? What output format do you want?

3. Start simple — begin with one workflow (like negative keyword mining), refine until reliable, then expand

4. Always approve — set up your skills to require confirmation before any mutation

The Bigger Picture

This is the future of performance marketing: not AI writing your ads, but AI operating your workflows.

"The marketer's job is evolving from executor to workflow designer," Austin noted. "You don't do the work — you design the logic, then review the AI's output."

The marketing teams that thrive will be the ones who can encode their expertise into reusable skills — then let AI handle the execution.

That's exactly what Austin built. And now you can too.