December 2025 · 6 minute read
I ran a blind A/B test between Gemini 3 Pro and GPT 5.2 on my slash command converter. Same prompt, same input. One gave me 60 lines I could use. The other gave me 250 lines I had to rewrite. The difference wasn’t the prompt—it was the training.November 2025 · 6 minute read
We’ve evolved from slash commands to MCP servers to skills, but our configs haven’t. That prompt from the Claude 3.7 era? It’s still running. That’s not legacy—it’s debt.November 2025 · 9 minute read
I tested three AI coding agents on writing analysis—Claude, Gemini, and Codex. After fixing a configuration bug, the results changed dramatically. Here’s what happened.