GPT2GTM
· Ravi

The N8N + Claude Stack That Runs My Agency

Running a performance marketing agency means drowning in repetitive tasks. Client reports. Status updates. Data pulls. Campaign summaries. It never ends.

Six months ago, I decided to automate everything I possibly could. The result? An N8N + Claude stack that handles 80% of my agency’s operational work.

Here’s exactly how it works.

The Core Stack

  • N8N — Self-hosted workflow automation (like Zapier, but way more powerful)
  • Claude API — For anything requiring intelligence
  • Google Sheets — Client-facing data layer
  • Slack — Notification and approval workflows

Workflow 1: Automated Weekly Reports

Every Monday at 6 AM:

  1. N8N pulls data from Google Ads, Meta Ads, and Analytics APIs
  2. Data gets normalized into a standard schema
  3. Claude analyzes week-over-week changes and writes insights
  4. Report populates in a Google Slides template
  5. Draft appears in Slack for my review
  6. One click to approve and send to client

Time saved: 4 hours per client per week.

Workflow 2: Anomaly Detection

Runs every 6 hours:

  1. Pull campaign metrics from all platforms
  2. Compare against rolling 7-day averages
  3. Claude evaluates if changes are statistically significant
  4. If anomaly detected: Slack alert with context and suggested action

No more logging into 5 platforms to check if something’s broken.

Workflow 3: Client Question Handler

When a client emails a simple question:

  1. Email hits a parsing workflow
  2. Claude categorizes: data question, strategy question, or needs human
  3. Data questions get auto-answered with fresh numbers
  4. Draft response goes to Slack for approval
  5. One click to send

The Key Insight

The magic isn’t in any single workflow. It’s in building a system where AI handles the first draft of everything.

I’m not trying to remove humans from the loop. I’m trying to make sure humans only do the parts that require human judgment.

Claude writes the first draft. I edit and approve. Clients get faster responses. I get my weekends back.

Getting Started

If you want to build something similar:

  1. Start with your most repetitive task
  2. Map out every step manually
  3. Identify which steps need AI vs. simple logic
  4. Build in N8N with approval gates everywhere
  5. Trust but verify for 2 weeks before reducing oversight

The goal isn’t full automation. It’s intelligent augmentation.