Jonathon's AI Wiki

Ben's Five-Skill AIOS Setup — OS Setup → Operator → Optimizer → Team → MCP

8 min read

Source: raw/5_Skills_to_Build_an_AI_Operating_System_Like_The_1%_Full_Guide.md — YouTube tutorial zElKhlFkqU4. Creator self-identifies as Ben running “an AI agency where we implement AI operating systems for businesses.” Sells “AI accelerator” community + 1-on-1 tech help + consulting. Skills shipped as a downloadable plugin ZIP (link in description).

A bundled five-skill kit that walks a single user (and eventually a team) from zero to a working Obsidian-backed second brain that Claude Code and Cowork can pull from. Each skill is a numbered stage; the kit is designed so you run them in order. Distinct from the Nate Herk AIS-OS 3-skill bundle (/onboard, /audit, /level-up) in that Ben’s bundle is opinionated about Obsidian + Relay-plugin permissions + Railway-hosted MCP — Nate’s is tool-agnostic about persistence layer.

Key Takeaways

Architecture overview

  • Second brain = a folder on disk, visualized via Obsidian (a free overlay). The folder, not Obsidian, is the source of truth.
  • Folder structure templates: solopreneur/professional vs business-with-team. The “business” structure includes context/, daily/, projects/, intelligence/, resources/, skills/, departments/, teams/, onboarding/. The bundled OS setup skill picks one and scaffolds the directory tree + CLAUDE.md files at each level.
  • CLAUDE.md as the routing map. Each subfolder gets its own CLAUDE.md that tells Claude where to find which information — explicitly credited in the transcript to “Andre Karpathy … one of the leading AI researchers” (and is the Karpathy LLM-wiki pattern this vault implements; see _index).

The five skills

1. OS setup — initial scaffold

  • Run /os-setup after picking a target folder.
  • Asks user vault type (solopreneur / professional / business).
  • Generates folder structure + per-folder CLAUDE.md files.
  • Runs a 12-section context interview — biographical, company, market, team, infrastructure, ICP, etc. Each section has brain-dump + file-upload boxes. Recommendation in transcript: “order a six-pack and a pizza, sit down for hours, use Whisper Flow voice transcription, dump everything.” Output is structured by Claude into the populated folder.
  • Final step: connect the folder as an Obsidian vault. Apply the PLN theme for visual polish (optional).

2. OS operator — daily real-time context refresh

  • Run /os-operator to set up a scheduled task in Cowork that pulls real-time data from connectors (Fireflies meeting transcripts, Circle community threads, Slack DMs, Google Calendar, email — whatever the user has connected).
  • Each run produces a daily context brief in daily/<date>.md containing critical escalations, today’s calendar, team task lists, meeting transcripts, file updates.
  • Can DM your co-founder/team via Slack when something urgent appears.
  • Limitation: scheduled tasks only run while the Cowork desktop app is open. To run autonomously with the laptop closed, route through Claude Routines + the OS MCP skill (#5).

3. OS optimizer — context bloat cleanup

  • Run /os-optimizer weekly or bi-weekly to audit the entire second brain.
  • Combines multiple named frameworks from the transcript:
    • Anthropic’s architecture / CLAUDE.md / dream framework / managed-memory best practices
    • Caveman compression method
    • Chroma context-route method
    • Karpathy LLM-wiki (this vault’s pattern)
  • Sample output cited in the video: 1,700 files audited, 34 problems found, 32 fixed, health score 46 → 94.
  • Fixes: dedup + merge, stale-context detection, broken wiki-links, format cleanup, irrelevant data removal, frontmatter + tags additions, folder reorganization proposals, conflicting-information resolution. Optimizes CLAUDE.md + index files for token efficiency.
  • Maps directly to the Lint operation in this vault’s vault schema (see Lint section for the 17-check parallel implementation).
  • Anthropic’s announced dream feature (currently only in Managed Agents) is positioned as a future replacement, but for now this skill ships the open equivalent.

4. Team OS — multi-user sync + role permissions via Obsidian Relay

  • Run /team-os once the single-user vault is working (transcript recommends ≥2 weeks of solo use first).
  • Installs the Obsidian Relay plugin (community plugin) for real-time multi-user folder sync. Set up by creating a Relay server, sharing the key with team members.
  • Then swaps the upstream Relay plugin for “Bennai Relay” — Ben’s custom fork that adds role-based read/write permissions (member vs owner). Use case: strategy doc readable by everyone, editable only by founders.
  • Comparison table from the transcript on shared-vault options:
ApproachReal-time?Permissions?Token efficient?Verdict
GitHubNo (manual push/pull)YesYesWorkable, not real-time
Google Drive / Notion via MCPYesYesNo — MCP adds latency, token cost, accuracy lossAvoid at scale
Obsidian Sync (native)No (click-to-sync)NoYesWorkable but no perms
Obsidian Relay + Bennai RelayYesYes (role-based)YesRecommended

5. OS MCP — Railway-hosted MCP server for autonomous routines

  • Run /os-mcp to expose the second brain as a remote MCP server so Claude Routines (cloud-hosted, run without laptop open) and Managed Agents can read/write to it.
  • Pipeline: installs Railway (referenced as railway install via Cowork or Code tab), guides through Railway account + API token, deploys the second brain as a server with a public vault-MCP URL.
  • User adds the MCP URL as a custom remote MCP in Cowork → Customize → Connectors → Add Custom.
  • Then connects to the Relay account (email + password from #4) to authorize the server against the actual vault.
  • End state: the same prompt that ran as a scheduled task can now run as a Routine in the Claude Code tab — selects “Remote” execution mode, picks a schedule, lists the connectors (Fireflies, Google Calendar, email, second-brain MCP). Runs even when the laptop is closed.

Mindset commentary (recurring in the transcript)

  • Start simple, day-one is not perfection. Ben’s vault was 30-40 docs at start, hundreds at 6 weeks, thousands now.
  • Context compounds. The agent your team has after 6 months is far more powerful than day-one — the value is in the accumulated context layer.
  • Always select the second-brain folder in every new chat across providers (Cowork, Claude Code, Codex, etc.).
  • If Claude doesn’t pull context automatically, instruct it explicitly to scan the vault before answering.
  • Shared business AI fights “AI always agrees” syndrome — a referenced “institutional AI vs individual AI” article (link in resources) is the rationale for team-OS investment.

Open Questions

  • “Ben“‘s full name + channel — only first-name identified in transcript. Brand naming suggests “Ben AI” / “Bennai” (custom plugin name = Bennai Relay).
  • Bundled plugin ZIP URL — described as “first link in description”; not fetched as ai-research yet. Worth pulling once URL is known to confirm skill contents.
  • PLN Obsidian theme — visualizer theme recommended in the video. Look up community-plugins listing.
  • Caveman compression method + Chroma context-route method — referenced as named frameworks the optimizer skill uses. Sources not in the transcript. Caveman is plausibly the Caveman Claude Code plugin (concise-mode for token compression — see nine-claude-code-plugins-10x-faster); Chroma is plausibly Chroma’s open-source RAG context-routing paper. Verification recommended before citing.
  • “Institutional AI vs individual AI” article — referenced as resource link, source not in transcript. Worth fetching.

Try It

  1. Pick a target folder. Anywhere on disk. Don’t worry about getting the structure perfect upfront.
  2. Download Obsidian (free) — Mac or Windows. Apply the PLN theme for visual polish (optional).
  3. If using the bundled plugin: download the ZIP from the video’s first description link, then in Cowork → Customize → Plugins → Create plugin → Upload plugin → drop the ZIP.
  4. Run /os-setup. Pick solopreneur or business. Let the 12-section interview run end-to-end (block 2-3 hours; use Whisper Flow or similar voice-transcription for brain-dump speed).
  5. Connect the resulting folder as an Obsidian vault. Open Obsidian → Open folder as vault → select.
  6. Set up daily refresh with /os-operator. Choose your connectors (Fireflies, Slack, etc.). Approve the scheduled task. Leave Cowork open for now.
  7. Run /os-optimizer after a week. Inspect the health-score report. Note what it merged/cleaned/flagged — this is the lint output for your second brain.
  8. (Multi-user only) Add team sync with /team-os after at least 2 weeks of solo use. Configure roles (member / owner) for each shared folder.
  9. (Autonomy upgrade) Set up /os-mcp if you want the daily refresh to run with your laptop closed. Deploys via Railway → exposes vault as MCP → routine in Claude Code tab.
  10. Alternative: if you already follow Nate Herk’s AIS-OS 3-skill bundle, compare framings — Nate’s onboard/audit/level-up vs Ben’s setup/operator/optimizer/team/MCP. Pick the persistence + permission setup that fits your team size.

Graph View

Backlinks