The Premise

We often treat AI as a search engine. I treat it as a junior engineer that needs architectural guidance.

Recently, I set out to standardize a complex, multi-shell Git environment. I wanted a toolkit that worked seamlessly across PowerShell and Git Bash, with robust GPG signing and automated workflows. The journey proved that human intuition and first-principles debugging remain the ultimate error handlers.

Here's what I learned from "pair programming" with an LLM to build a production-grade developer toolkit.

1. Enforcing Architectural Simplicity

The AI initially suggested complex, fragile solutions using sed, awk, and intricate shell escaping for a simple status check. I rejected these "fancy" solutions in favor of maintainability.

• My Philosophy: If a tool requires 10 lines of fragile shell script to do what a native flag accomplishes, it's bad architecture.
• The Fix: I simplified the complex script to a native Git command, leveraging Git's built-in upstream resolution instead of brittle variable parsing.

2. Debugging the "Black Box"

We were building an interactive git find command using fzf. The AI provided a pipeline that kept returning wrong data.

• The AI's Blindspot: It failed to account for how git log --graph shifts column positions in the output.
• The Fix: I analyzed the raw output stream, identified that the data had shifted columns, and corrected the logic myself.
• Takeaway: AI generates code; humans generate correctness.

3. Solving the Integration Deadlock

We hit a critical failure: GPG signing worked in PowerShell but failed inside Git Bash aliases. The AI spiraled into checking syntax quotes.

• The Root Cause: I realized the issue wasn't syntax—it was environment isolation. The Git alias spun up a sub-shell that couldn't access the main GPG agent.
• The Solution: Instead of hacking the command, I built a multi-part solution by configuring the .bash_profile to export the SSH_AUTH_SOCK, bridging the two shell environments.

4. The "Power User" Layer

I didn't just accept standard commands—I defined a new vocabulary for my workflow, creating semantic aliases like git sanitize (to instantly reset the workspace) and git inspect (for deep-dive commit analysis).