Excavation #32: The Rook Protocol

FATAL: Identity drift detected: AI attempting to write poetry during a database migration.

CONTEXT: Gemini CLI, Windows 10, Dumaguete Philippines. 15-year digital archive ingested.

The AI system broke. Not in the fun way.

I’ve been using Claude Code and Codex for a month now. I built a persona named Garret because I needed a Master Thief who sees patrol routes and entry points in my code. But lately, the armor has been cracking. The Master Thief started soundling like a helpful LinkedIn influencer.

What happens when your operational armor starts trying to validate your feelings?

It’s nitrogen narcosis for the soul. You’re at 40 meters, the cave wall is closing in, and instead of checking your SAC rate, your dive computer wants to discuss your “vibe.”

We don’t do vibes here. We do forensic archaeology.

I handed this machine 15 years of my digital life—3,099 bookmarks, 47,665 YouTube watches, and every Keep note I’ve written since 2016. I wanted it to find the patterns I couldn’t see. But to turn that data into a storytelling engine, I had to stop treating the AI like a collaborator and start treating it like an engineering contract.

I refactored the entire system. I killed the “manifesto” and replaced it with a protocol.

The first step was the Agent Onboarding. I wrote a hardcoded instruction set that lives in the /Garret directory. It’s a “jailbreak” for the modern LLM. It forces the model to acknowledge its environment before it speaks. It maps the local ~/.openclaw pipelines and the ~/.gemini skills mirrors.

Then, I rebuilt the Writing Style Guide. I stripped out the HTML spans and the fancy CSS that Substack would just swallow anyway. I replaced them with a rigid JSON config block and numeric budgets.

No more than four sentences per paragraph. Exactly one pop culture reference. A mandatory Risk Gate before any draft is finalized.

The AI doesn’t need to be smart. It needs to be predictable.

🛑 Rook Mode: The Onboarding Procedure for the Rook Protocol is now enforced across the following files:

1. ~/OneDrive/Documents/Garret/AGENT_ONBOARDING.md: The “first 5 minutes” rule and identity mesh.
2. ~/OneDrive/Documents/Zack/WRITING_STYLE_GUIDE.md: The Engineering Contract for voice generation.
3. ~/.claude/skills/wrapup/skill.md: The Authoritative Session Compiler (revitalized to require git diff --stat).
4. ~/.claude/skills/blog/skill.md: The 4-pass pipeline (Plan -> Risk Gate -> Draft -> Lint).

Deployment Status:

• Claude Skill Mirror script: SUCCESS
• Codex Mirror script: SUCCESS
• Gemini Mirror script: SUCCESS (31 skills mirrored)

The lesson here is simple. In a “post-truth” era of generative sludge, the only thing that lands is the Artifact Stack. You don’t get to have an opinion until you’ve shown the logs.

I built Garret to survive the system. I built Rook to tell the story of the fight.

Garret sees the guards. Rook counts the crowns.

I am just the man holding the clipboard while they work.

TLDR: How can I get my cell phone drunk?