The Journey
The Problem: 10 Years of Digital Hoarding
After a decade at SAP, I was drowning in resources. My solution? Browser bookmarks that grew into an unmanageable mess. When Chrome introduced Tab Groups, I thought I'd found salvation—until I realized I was maintaining 200+ tabs across multiple groups. The anxiety from my digital FOMO was worse than ever.
Back to Basics (With a Twist)
I returned to curated bookmarks but with a new goal: create a "PM Toolbox" that I could easily maintain and share with colleagues across the organization. I built a clean bookmark folder and mirrored it on a wiki page.
The Maintenance Nightmare
Within weeks, reality hit: my bookmark folder evolved constantly, requiring manual wiki updates. Chrome's bookmark export format (the ancient Netscape standard) made extracting subsets error-prone and frustrating. I needed automation.
First Attempt: GenAI Fails the Coding Test
I tried using Claude 4.5 Sonnet through GenAI-XL. Despite multiple attempts, it couldn't handle Chrome's non-standard HTML format. It kept reaching for HTML parsing libraries that failed, then suggested an obscure Netscape parser that violated SAP standards. Dead end.
Second Attempt: GitHub Copilot Nails It
The next day, I switched to VSCode with GitHub Copilot. Here's what changed:
- I wrote clear requirements in a `.github/copilot-instructions.md` file, specifically calling out the Netscape format complexity
- Copilot enhanced my requirements, suggesting testing coverage and documentation guidelines (80% of the final instructions came from Copilot)
- I switched to Agent mode and prompted: "Please create the missing files. Do not reference public code in your response." (Critical for enterprise compliance)
The result? A complete, working script—instantly.
But it got better. When I asked for a second script to generate Confluence tables, Copilot:
- Created the implementation
- Built the test suite when I requested it
- Ran the tests autonomously, discovered bugs, and fixed them—all without me asking
I was genuinely amazed.
Key Takeaways
1. Match the Tool to the Task
- General AI (like Claude): Excellent for brainstorming, requirements refinement, and PM tasks
- Specialized coding AI (like GitHub Copilot): Purpose-built for implementation and testing
2. Clear Requirements Are Non-Negotiable
Both attempts used AI, but only the second succeeded because I provided structured, detailed requirements upfront.
3. Embrace Test-Driven Development
Even with AI assistance, automated testing is essential. Copilot's self-correction only worked because tests exposed the bugs immediately.
4. Simple Solutions Often Win
Sometimes the answer isn't a shinier tool—it's going back to basics (curated bookmarks) with smart automation layered on top.
The Bottom Line
AI coding assistants aren't magic—they're power tools. Like any tool, they require the right context, clear instructions, and quality controls. When you provide those, the results can be transformative.
For PMs and non-developers: you don't need to be a coder to automate your work. You just need to understand your requirements so you can clearly write them down and choose the AI tool designed for the job.
References
You can find the scripts in question here 
Bluesky