I Turned Claude Opus 4.8 Into My Entire AI Operating System
Source: YouTube: Nate Herk | AI Automation
Summary
Nate Herk walks through his “Herk 2” AI Operating System built around Claude Code + Opus 4.8. The key insight: context is king, not the model. Everyone has access to the same models — the moat is your knowledge architecture.
The Default Shift
Reach for Claude Code first before Chrome or any other app. If the use case is brainstorming, writing, thinking — not just coding — use Claude Code anyway. This builds up your context continuously and eliminates context switching.
“AI isn’t king. Opus 4.8 benchmarks — who cares? Everyone has access to the same models. If the AI is king, wouldn’t everybody be king? Context is king. “
The Four C’s Framework
| Layer | What it means |
|---|---|
| Context | Does it know your business? Open a fresh session → it should be able to describe what you do, who works there |
| Connections | What can it actually touch? Calendar, tasks, Slack, email, QuickBooks, CRM — via APIs / MCP servers |
| Capabilities | How does it do work? Skills = instruction files (writing guides, frameworks, SOPs) |
| Cadence | Things that happen while laptop is closed — scheduled automations |
Each layer depends on the previous one.
The Three M’s (Mindset)
- Mindset — Shift to “do everything through the AIOS”
- Method — The 4 C’s architecture
- Machine — The tools (Claude Code, Codex, etc.)
Building Skills
Two approaches:
- Build forward — Tell Claude Code “I want a skill for X” and iterate (~50 tries sometimes)
- Reverse engineer — Complete a task end-to-end, then have Claude Code look back at the conversation and extract the skill from what worked
Skills can be dead simple — even just a saved prompt you type repeatedly. Example: session handoff skill (summarize context, open decisions, key files — so you can clear context and pick up fresh).
The Bike Method (Trust Building)
Don’t hand the keys over immediately. Like teaching a kid to ride a bike:
- Walk with them, hold the handles
- Training wheels
- Let them ride down the street while you watch
- Eventually full autonomy
Each iteration builds trust. Instructions ≠ capabilities — if an email-sending tool is connected, the agent can use it even if you said “don’t.”
Real cautionary tale: An AI agent on his team proactively picked up a task and sent 3 promotional emails to 150K+ inboxes that weren’t supposed to go out. Permission layers matter.
Organization
- Everything is files and folders — tool agnostic (works with Claude Code, Codex, OpenClaude)
agents/,skills/,other-worlds/(separate Claude Code projects)CLAUDE.mdchanges almost daily — don’t stress about getting it perfect- Quarterly reprioritization = folder restructuring is normal
No Dashboard Needed
Nate doesn’t use a visual dashboard (Obsidian knowledge graph etc.). He works directly in the terminal. Productivity isn’t hours worked — it’s “did I move the needle closer to my goal.”
Key Quote
“You can outsource your thinking, but you cannot outsource your understanding.”
Nate Herk runs an AI automation agency. Free GitHub repo + 3-hour course on building your own AIOS available in his free School community.