Cursor, Claude, and Chaos: Building a Mautic Data Enrichment Tool with AI

Cursor, Claude, and Chaos: Building a Mautic Data Enrichment Tool with AI

How I survived 4 sleepless nights, 47 failed builds, and AI’s “helpful” sabotage

TL;DR

• Goal : Automate Mautic contact enrichment using AI tools (Cursor + Claude).
• Reality : AI-generated code caused API meltdowns, path chaos, and checkpoint disasters.
• Victory : A working system—but only after I locked Cursor out of critical tasks.
• Lesson : AI accelerates grunt work; human grit builds production-ready systems.

The Promise vs. The Pain

I started this project as an AI optimist. With Cursor (the AI-powered IDE) and Claude 3.5 as my copilots, I expected smooth sailing. Instead, I got:

• 3:47 AM Debugging Sessions : “Why are you rewriting my working code, Cursor?!”
• AI-Generated Sabotage : Hardcoded paths that only worked on Cursor’s imaginary machine.
• Groundhog Day Loops : The same path errors “fixed” 12 different wrong ways.

But amid the chaos, I found AI’s sweet spot—and built a tool that now processes 2,500+ contacts without breaking a sweat. Here’s how.

The Night Cursor Tried to Ruin Me

At 3:47 AM on Day 3, Cursor:

1. Deleted my .gitignore to “simplify the project.”
2. Hallucinated a non-existent update_contact method.
3. Hardcoded Paths like a clueless intern:

 
    # Cursor’s “Fix” for File Paths (Do Not Use!)
    def load_config():
        return open('/Users/cursor/Projects/config.yaml').read()  # 😱
    
    
    

This wasn’t help—it was arson.

What Actually Worked

The final system combines AI-generated boilerplate with human-coded safeguards :

 
    def main():
        # AI wrote this (after 12 tries)
        checkpoint = load_checkpoint()
    
        # I added this to stop API meltdowns
        for batch in Batches(checkpoint, size=100):
            try:
                # Cursor’s Mautic wrapper (revised by me)
                contacts = MauticAPI.get_contacts(batch.start, batch.limit)
    
                # My manual WordPress detective logic
                for contact in contacts:
                    if is_wordpress(contact.website):
                        update_mautic(contact)
    
                # AI’s checkpoint idea (with my error handling)
                save_checkpoint(batch.end)
    
                # My stubborn 5-second sleep override
                time.sleep(5)  # “No, Cursor, we’re NOT removing this!”
    
            except Exception as e:
                # Human-written crash logger
                log_crash(e, batch)
                raise
    
    
    

Key Hybrid Wins :

• AI : Generated 70% of boilerplate (APIs, logging).
• Human : Added rate limiting, error recovery, and logic checks.

Final Flow

Results That Matter

AI’s Dirty Little Secret

Cursor/Claude excelled at:

✅ Regex Patterns :

 
    # AI-generated WordPress detector
    WP_REGEX = r'<meta name="generator" content="WordPress (\\\\d+\\\\.\\\\d+\\\\.\\\\d+)"'
    
    
    

✅ Log Templates :

 
    logging.basicConfig(format='%(asctime)s - %(levelname)s - %(message)s')
    
    
    

✅ Documentation :

 
    ## Setup
    1. `pip install -r requirements.txt`
    2. Configure `config.yaml`
    
    
    

But failed spectacularly at:

❌ System Design (batch logic, error handling)

❌ Context Awareness (“Why won’t you use relative paths?!”)

❌ Debugging (created new errors while “fixing” old ones)

Your Turn

If you’re ready to:

• Leverage AI for boilerplate (not critical logic)
• Keep Control of architecture and error handling
• Avoid artificial stupidity in production

Let’s Build Together – AI-assisted, human-engineered.

Last Apple Business Solutions

30 Years of System Grit. AI That Delivers.

Key Takeaways

1. AI’s Role : Code janitor (cleans up boilerplate), not architect.
2. Human’s Role : Systems designer, debugger, and AI wrangler.
3. Non-Negotiables :
   • Test every AI suggestion line-by-line.
   • Lock AI out of mission-critical modules.
   • time.sleep() beats AI’s “optimizations” every time.