Watch on YouTube: https://www.youtube.com/watch?v=bCljOfCH8Ms
Channel: Nate Herk | AI Automation
Building a Personal AI Operating System with Claude Code: A Complete Walkthrough
Nate Herk’s two-hour course on building and selling AI Operating Systems (AIOS) using Claude Code in VS Code is one of the most practical deep-dives available on agent-based productivity. Rather than chasing flashy demos, the course centers on a repeatable framework: a local, version-controlled system that sits between you and your work, letting agents handle routine tasks without constant context-switching. The entire template is free on GitHub (accessed via Nate’s School community), but you’ll need a paid Claude account to run Claude Code.
The Core Frameworks: Three M’s and Four C’s
Herk structures the build around two mental models. The Three M’s define how you approach the work:
- Mindset: Default shift to AI — before any task, ask how AI could handle 30% of it. Break work into function trees. Treat AI as a mentor, not a vending machine.
- Method: Assign a leverage percentage to each task. Automate one chunk at a time. Expect a 20% productivity dip before a 50%+ gain.
- Machine: The technical implementation — skills, connections, loops, and the file structure that makes it durable.
The Four C’s dictate build order:
- Context: What the AI knows about you, your business, and your preferences.
- Connections: Data access via API or CLI (not MCP) for token efficiency.
- Capabilities: Reusable skills and SOPs encoded as markdown prompts.
- Cadence: Autonomous routines — daily plans, weekly audits, recurring reviews — powered by Claude Code’s new loop/cron feature.
Step-by-Step Walkthrough
1. Scaffold the Repository
Clone the free GitHub template into a local folder, open it in VS Code, and initialize Claude Code. The repo ships with a .claude/skills directory containing three foundational skills: onboard, audit, and levelup, plus a CLAUDE.md master prompt that governs agent behavior.
2. Run the Onboarding Skill
Type /onboard in Claude Code. The skill runs a seven-question interview (voice dictation via Glide is recommended) and scaffolds your day-one context files: about_me.md, about_business.md, priorities.md, and writing_style.md. These become the grounding layer every subsequent skill references.
3. Map Your Tier-1 Domains
Identify the seven core domains that drive your work: revenue, customer, calendar, communications, tasks, meetings, and knowledge. Each domain will eventually get its own connection and capability layer.
4. Wire Connections via API (Not MCP)
Herk emphasizes direct API integration over MCP for token efficiency. For each tool — ClickUp, Google Workspace, Fireflies, QuickBooks — create a scoped service account (e.g., up_ai), generate a token, and store it in .env. Ask Claude to research the API docs and produce a reference markdown file in connections/. The Google Workspace CLI (gws) is a standout: a single binary with 100+ built-in recipes covering Drive, Gmail, Calendar, Docs, Sheets, and Slides.
5. Audit and Level Up
Run /audit to score your Four C’s maturity. Then run /levelup, which asks five diagnostic questions to surface high-leverage automation opportunities. The output is a prioritized backlog of skills to build.
6. Build a Karpathy-Style LLM Wiki in Obsidian
Adopt Andrej Karpathy’s knowledge pattern: drop source documents into a raw/ folder, then run a prompt that transforms them into an auto-organized wiki/ with an index, changelog, and interlinked markdown nodes. Query this graph in Obsidian (free for personal use, plus the Web Clipper for ingestion). One X user reported a 95% token reduction when querying the wiki versus scattered files.
7. Establish Cadence with Loops and Crons
Claude Code’s loop/cron feature (session-scoped, 3-day max expiry) lets you schedule recurring tasks in natural language: every 20 min run review_pr skill. Recommended baseline: a morning plan my day loop, an evening review, and a Friday audit + skill-usage review.
Tools, Numbers, and Costs
- Claude Pro: $17/mo (annual) or $20/mo — required for Claude Code.
- VS Code + Claude Code extension: Free.
- GitHub template: Free via Nate’s School community (Classroom → YouTube Resources).
- Google Workspace CLI (gws): Free, single binary, JSON-first, auto-discovery.
- Obsidian: Free for personal use; Web Clipper extension for ingestion.
- Glide (voice dictation): Affiliate link in description; Herk switched from Whisper Flow.
- ClickUp API: Personal token + team ID stored in
.env. - Loop/cron limits: 3-day expiry, session-scoped; commands:
cron create/list/delete. - Token reduction claim: 95% drop when querying Karpathy wiki vs. scattered files (per X user, unverified).
Caveats and Hype Check
Nate claims a $100k/mo agency exit and a 350k-member community — plausible given his output but unverified. Heavy affiliate push for Glide with an “official team member” disclosure. AIOS results are personal anecdotes; no independent benchmarks. The template is free but requires paid Claude Code. Revenue claims serve as social proof for course authority.
In practice, the system’s value depends on how deeply you live in it. Herk’s own rule applies: “You can’t scale a system you haven’t lived in yourself.” Expect a learning curve — the 20% dip is real — and budget time for API debugging and skill iteration.
Concrete Next Actions
- Join Nate’s free School community (link in video description) and download the AIOS GitHub repo.
- Install VS Code + Claude Code extension; log in with a paid Claude account.
- Create a local AIOS folder, open in VS Code, clone the repo via Claude Code.
- Run
/onboardskill; answer the 7 questions (voice dictation via Glide recommended). - Map your 7 tier-1 domains: revenue, customer, calendar, comms, tasks, meetings, knowledge.
- Wire your first connection (e.g., ClickUp): create service account, get API token, store in
.env, ask Claude to research API docs and build reference markdown. - Run
/auditto score your 4 C’s; run/levelupanswering 5 questions to find automation opportunities. - Install Google Workspace CLI (
gws) if in Google ecosystem; test with “create YouTube resource guide.” - Set up Karpathy LLM Wiki: create Obsidian vault, paste Karpathy prompt + Nate’s wrapper, ingest first source via Web Clipper.
- Establish daily loop: morning “plan my day,” evening review; weekly: Friday audit + skill usage review.
The AIOS approach isn’t a magic bullet — it’s a disciplined engineering practice. Start small, automate one chunk at a time, and let the system earn its keep before expanding. Boring, deterministic workflows still beat fragile agent chains nine times out of ten.