- Entries
- 03 published
- Updated
- Jul 2026
- Archive
- All posts
I Automated YouTube Video Triage, Then Found Out I Was Optimizing the Wrong Thing
I built a tool that flags my weak YouTube videos and rewrites their titles. It felt smart. Then I had OpenAI's Codex red-team the design and checked every claim against YouTube's own guidance. Rewriting metadata turned out to be the wrong fix, and my success metric was measuring noise.
Gemma vs Claude for SEO Scoring: What the 26-Point Gap Reveals
Same posts, same prompt, two AI models, a 26-point gap. I built dual AI scoring into Cerebro to compare Gemma 4 and Claude Haiku side by side, and the disagreement says more about modern SEO than most checklists do.
Stopping Claude Code From Leaking Secrets
Claude Code leaked a real token fragment into my chat session during a routine operation. After two incidents and a painful key-rotation cycle, I built a two-hook safety layer so Bash output gets redacted before the model sees it.