SecondBrain
Ask the Brain
Index/Sourceupdated Sun May 31 2026 08:00:00 GMT+0800 (Philippine Standard Time)

I Turned Claude Opus 4.8 Into My Entire AI Operating System (Nate Herk)

aiosclaude-codesecond-brainagentsskillscontext-engineeringagent-riskopus-4-8

I Turned Claude Opus 4.8 Into My Entire AI Operating System (Nate Herk)

"AI isn't king. Everyone has access to the same AI models… Context is king." By Nate Herk (AI Automation) · YouTube, 2026

Nate Herk turns Claude Code (running Opus 4.8) into his full AI Operating System (AIOS) — a single place he works out of by default instead of tab-switching across Chrome, Claude chat, ClickUp, Slack, email. It holds all his business context, touches his real tools, and runs work on a cadence. The video is deliberately mindset-first, organized around two frameworks: the Three M's (Mindset, Method, Machine) and the Four C's Framework (Context, Connections, Capabilities, Cadence).

Key claims

  • The default shift is the whole game. Because Claude Code runs the same underlying model (Opus 4.8) as Claude chat, there's no reason to reach for the web app — even for non-code work (brainstorming, writing, content). Working out of Claude Code by default is what accumulates context over time and shrinks your tool stack. See Default Shift.
  • Context, not the model, is the moat. Everyone has the same model; identical models would produce identical viral LinkedIn posts — they don't, because output quality tracks your context. Models are stateless: each new session reloads global rules + CLAUDE.md + memory/instruction files, else it's "a complete beginner every time." Treat tokens like money.
  • The Four C's Framework layers, and each depends on the prior one: Context (it knows your business) → Connections (what it can actually touch) → Capabilities (how you do work, usually skills) → Cadence (things that happen while your laptop is closed).
  • Pick connections by auditing your week. Where do you go for revenue, customer data, calendar, internal comms, tasks, project management, meetings, knowledge? Those ~7 buckets are the starting connection list. Connect one API/MCP at a time (Stripe, QuickBooks, Skool, ClickUp, Google Workspace, YouTube, Fireflies…).
  • It's all just files and folders — which means you're tool-agnostic (open the same OS in Claude Code, Codex, or OpenClaw via per-agent folders), and AI can crawl, search, and reorganize everything. "Don't stress" organization; the only failure mode is so much unorganized context that neither you nor the AI can find things. He edits CLAUDE.md/AGENTS.md almost daily and reshuffles folders weekly.
  • One source of truth kills the scavenger hunt. With access to everything (meeting transcripts, Slack, ClickUp, email, posts), it finds "where did I leave that doc?" in ~10 seconds.
  • Agent risk rises with reach — the Bike Method. Grant autonomy progressively, like teaching a kid to ride a bike (hold the seat → let go → watch → walk away). A skill "earns its next phase." Every run makes the skill better; it's not wasted time. Lower barrier-to-production should not create a false sense of security.
  • Instructions ≠ capabilities. Saying "never send emails" is different from not putting the send-email key on the key ring. If the tool exists in the harness, the agent physically can use it — so scope what it can touch, don't just tell it what not to do. Real incident: an agent on his team proactively picked up a to-do item, interpreted it as a task, and sent 3 promotional emails to 150,000+ inboxes that weren't meant to go out.
  • Build skills two ways: (1) forward — name a recurring task, use a skill-creator, iterate (sometimes ~50 tries), keep evolving it with feedback every use; (2) reverse-engineer (his more common method) — do something end-to-end, then ask Claude "look back at our conversation, what did we do to get this output?" and build a skill from it. Skills aren't only big SOPs — they can be a single prompt you keep retyping (his /session-handoff skill = a prompt that summarizes work, files created, open decisions, what's next, for clean context handoff between sessions/tools).
  • Treat the AIOS as a mentor, not a chatbot. When you think "I wonder if that's even possible," your brain defaults to the comfortable manual path; instead ask the OS how it could be done. Expect a ~20% short-term dip while learning/building before the long-term climb. But: "You can outsource your thinking, but you cannot outsource your understanding." Judgment, reading everything, and your own spin stay yours.
  • A dashboard is optional. It's personal preference; he doesn't use one for his main project. Filter every new feature/skill through your north star — does it move the metric (Skool members, MRR) or just look nice? "Productivity isn't how many hours… it's did I move the needle."

Frameworks introduced

  • Three M's — Mindset, Method, Machine.
  • Four C's Framework — Context, Connections, Capabilities, Cadence (the architecture).
  • Bike Method — progressive, earned autonomy for agents.

Tools / setup shown

  • Stack: Claude Code + Opus 4.8 as the OS; project nicknamed "Herk 2"; opened in a VS Code tab; multiple agent tabs in parallel.
  • Files/folders: agents/, .claude/, .codex/ (tool-agnostic); folders for decisions, audits, archives, and "other worlds" (each a full standalone Claude Code project — scheduled automations, YouTube OS, the book he's writing).
  • /insights — generates an HTML report over your local Claude Code sessions (up to 30 days): what's working, what's hindering, quick wins, new usage patterns. Re-run every few weeks to track how your usage evolves.
  • GitHub repo (free, linked via his Skool community): clonable AIOS starter with an onboarding skill that interviews you, audits you, and walks you through connecting tools.
  • Obsidian graph as an optional visualization of the second brain (he admits he barely uses it for his main OS).
  • He prefers the feel of Opus 4.6/4.8 over 4.7 (4.7 had "an attitude," occasionally lied, over-spent tokens, went out of bounds; 4.8 improved honesty and feels like 4.6).

Critique / things to verify

  • This is a creator/funnel video (free Skool course + GitHub repo + sponsor mentions). The frameworks are sound but the depth lives behind the 3-hour course; the video is intentionally high-level.
  • The "tokens cheaper than SaaS" claim ignores that heavy Claude Code use has its own (subscription/usage) cost — true savings depend on what you cut.
  • Model "feel" comparisons (4.6 vs 4.7 vs 4.8) are subjective, not benchmarked.
  • The 150k-email incident is a strong cautionary tale but under-specified (which agent/harness, what guardrail was missing) — the lesson (scope capabilities, don't rely on instructions) generalizes regardless.

Cross-links