I Turned Claude Fable Into The Ultimate Second Brain (Nate Herk)

Nate Herk (AI Automation)'s full walkthrough of his "Herk 2" AIOS one day into the Fable 5 launch (June 9, 2026) — the most operational tour of his setup yet, structured as the Four C's Framework with a sharpened two-layer split: Context + Connections = the second brain; Capabilities + Cadence = the AIOS built on top of it.

Key claims

1. An OS doesn't start with architecture — it starts with a default. Close the other tabs; do everything through one harness (Default Shift). The gut check: ask it about your business — does the answer sound like a stranger or a co-founder?
2. CLAUDE.md is a router, not a knowledge dump. Context is a routing tree: the file points to where rules, skills, wikis, and projects live rather than containing them. He coins Architecture Engineering as the emerging discipline next to context/prompt/harness engineering; pulse check = "could I manually drill to the file, and can my agent?" His large setup still opens at ~40k tokens of context.
3. "Other worlds" folder — he moved his formerly separate repos (YouTube OS, book, website) inside the main project so one agent can cd between them and one GitHub push syncs everything. (The structural twin of this machine's ~/worlds.md hub — see AI Operating System (AIOS).)
4. Connections = the live-data tier. Audit your week (revenue, customers, calendar, comms, tasks, meetings, knowledge) and wire tier-1 apps first, via APIs/CLIs over MCP ("more control, cheaper") — consistent with his CLI vs API vs MCP position. Static data goes stale visibly (his AIOS reported 620k subscribers when he's at ~800k).
5. Skills compound through a feedback ritual: every use ends with "here's what I liked / didn't — update the skill." A skill can be just a prompt you keep retyping; his "grill me" interview skill (adapted from Matt PCO) extracts knowledge from his head into the AIOS — he used it to plan this very video. See Skills (Claude Code).
6. Cadence is earned, and the permission layer is keys, not prompts. "A prompt is never a permission layer… if it could send an email, it might." Retells the 150k-recipient discount-email incident as the case for scoped API keys — read-only transcripts, no send capability (Capabilities vs Instructions (Agent Keys), Bike Method).
7. Fable 5 hands-on: ate his $200/mo Max 5-hour session limit in ~1 hour of stress-testing; "it just gets it" feel (citing Andrej Karpathy's tweet — whom he says has joined Anthropic); safeguards "a little too trigger-happy for launch." One-shot showcases: an auto-generated channel-history video and a 21-minute one-shot concept-relationship web app built from his transcripts. Delegate parallel work to Sonnet/Haiku to manage cost (Model Routing).
8. You're not building a Claude Code AIOS — folders, markdown, skills, and routing logic are tool-agnostic IP; he keeps parallel claude.md/agents.md configs and runs the same OS in Codex.

Usage tips (his top three)

• Thought partner with devil's-advocate sub-agents debating perspectives — counter the sycophancy tendency.
• Have it interview you (grill-me) rather than writing context docs by hand.
• Make it verify its own work — dynamic workflows, Playwright click-throughs, persona checks ("would a beginner understand this?"); raises first-pass quality from ~70% to ~92%.

Vault notes

• Strongest practical articulation yet of the AIOS thread; directly validates this machine's ~/worlds.md + skills + cadence setup (see Turning This Mac Into an AI Operating System).
• The Karpathy-joined-Anthropic claim is secondhand via this video; not independently verified here.