Manifesto
Why a Personal OS?
Because AI literacy is a commodity. Personal infrastructure is not. The leverage isn't in knowing prompts — it's in running a system that survives your bad days.
The 4 levels of AI usage
-
Level 1 — Chatbot user
One-off questions. Copy-paste answers. No memory, no leverage. This is where most people stop.
-
Level 2 — Power prompter
Longer conversations, role-play, custom instructions. Better output, same manual loop.
-
Level 3 — Tool builder
Scripts, automations, MCPs, custom GPTs. Real productivity, but fragile and scattered.
-
Level 4 — OS operator
Memory, rules, capability registry, scheduled tasks, safety constraints — running as one system. AI is a teammate with context, not a tool you launch.
Core beliefs
- Specific knowledge over general skill. The moat is in how you configure the system for your life and work, not the model.
- Capability-first, not prompt-first. Never write a script for something Claude already does natively.
- Memory before intelligence. A smart assistant with no memory is a stranger every morning.
- Rules over reminders. Hard constraints beat polite requests. The system should fail safely.
- Ship, don't theorize. Every rule was earned in production. The OS is a log of lessons.
The Level4 thesis
Functional experts will out-automate you in their own domain. A senior ops person with Claude will beat you at ops. A senior CS person will beat you at CS. Competing on domain knowledge is a losing game.
What doesn't scale with headcount is systems thinking applied to your own life. A Personal OS is the architecture you carry across roles, companies, and decades. It compounds. Prompts don't.
Level4-OS is my running proof. Nine live subsystems, a memory layer, a capability registry, hard safety rules, and a weekly self-audit loop. It is the artifact on this site.
Architecture
Nine live subsystems. One operating system.
Level4-OS is organized the way an OS is: kernel, memory, rules, capabilities, observability. Every subsystem has a clear job. Every subsystem is documented.
01 · Kernel
Claude + Cowork
Execution substrate. Where every request lands, reasons, and acts.
02 · Memory
Identity + Context + Learnings
Persistent files that give Claude continuity across every session.
03 · Rules
Capability-first, Safety, Output
Hard constraints that govern tool selection, cost, and destructive actions.
04 · Capability Registry
CLIs · MCPs · Plugins · Scripts
Typed map of every tool, its cost, and when to use it.
05 · Scheduler
Cowork Scheduled Tasks
Recurring jobs — health checks, outcome refreshes, learnings consolidation.
06 · Observability
Health status + run logs
Every scheduled run produces an artifact. Nothing is opaque.
07 · Router
Project + thread dispatch
Which project context each request belongs to. Prevents cross-context leakage.
08 · Command Center
Unified dashboard
Single HTML artifact showing systems, business IP, observability, router.
09 · Public Layer
Site + Starter Kit
The credibility artifact you are reading, and the lead magnet it feeds.
Memory layer
Three files do most of the work: identity.md (who I am), context.md (what I'm doing, why, who I work with), and learnings.md (what I've learned the hard way). Loaded at the start of every session. Updated at the end.
Capability tiers
Strict order: native shell → official CLI → Cowork native → official MCP → plugin → unofficial → script → manual. Token cost, stability, and recovery time all slope the same way. The rule removes decision fatigue.
Capabilities
The stack, top to bottom.
Most AI stacks I see are a pile of tools. A Personal OS has tiers. Lower tiers are cheaper, more stable, and always preferred.
Tier 1 — Native shell & bash
Zero token taxThe operating system itself. File operations, git, curl, scheduled tasks, pipes. Most stable. Never loads schema tokens.
Tier 2 — Official CLIs
StableVendor-built command-line tools. supabase, vercel, gh, qpdf, sentry-cli. Battle-tested. Preferred over MCPs when both exist.
Tier 3 — Cowork native
IntegratedScheduled Tasks, Claude in Chrome, built-in skills. First-party. No extra install, no extra auth.
Tier 4 — Official MCP connectors
When CLI is missingNotion, Granola, Apollo, Common Room, Ahrefs, Gamma, Canva, Figma. Chosen only when no CLI exists or auth would be fragile.
Tier 5 — Plugins
PackagedBundled skills + connectors for a role (sales, marketing, ops). Good for batch workflows, but expensive in token load.
Tier 6 — Scripts & manual
Last resortCustom scripts or manual steps. Only when no higher tier fits. Every script is tech debt waiting to break.
The Capability-First rule
Before suggesting any tool, I walk the registry top-down. If a native command does the job, MCPs don't get loaded. If an official CLI exists, I don't install a third-party wrapper. This one rule removed ~40% of the token waste and ~80% of the flakiness in my stack. It's the single highest-leverage rule in Level4-OS.
Outcomes
What the system is producing.
Last refreshed
2026-04-19
Refreshed by hand from the private Level4-OS. Everything below is sanitized — aggregated counts, categories, and lessons, never client names or private data.
Scheduled automations live
14
across memory, outreach, and observability
Hours reclaimed / week
~12
rolling 4-week average
Recent learnings (sanitized)
-
WEEK OF 2026-04-13
Stability > cleverness in scheduled tasks.
A flaky MCP took down a weekly sync for three consecutive runs before I noticed. Replaced it with a two-line bash + official CLI. No failure since. Lesson: MCPs pay token rent; make them earn it.
-
WEEK OF 2026-04-06
Separate the memory file per project.
Merging two client contexts into one context.md leaked private vocabulary across domains. Split them, gated by the router. No cross-context bleed since.
-
WEEK OF 2026-03-30
Hard rules beat soft reminders.
"Please ask before destructive actions" was ignored twice. Converted to a hard constraint: no destructive command without explicit token in the prompt. Zero incidents since.
System health: green. No silent failures in the last 14 days.
Starter Kit
Level4 Starter.
The skeleton of a Personal OS, stripped down to the parts you can set up in 30 minutes. Templates teach structure. The PDF teaches philosophy. Together they get you from Level 2 to Level 3.
Get the Starter Kit
Free. Email-gated so I can send a 5-part setup nurture. Unsubscribe any time.
You'll receive a confirmation email to verify your address (double opt-in). The Starter Kit download link arrives after you confirm.
What's inside
4-page PDF wrapper
- · Why a Personal OS
- · The Level4 architecture at a glance
- · 30-min setup guide
- · DIY → full build-out upgrade path
Template files (zip)
- · CLAUDE.md with placeholders
- · Memory layer scaffolding
- · Safety + capability rule copies
- · Blank learnings & capability registries
FAQ
Do I need Claude Pro / Cowork to use the Starter? ›
You need a Claude plan that supports projects and memory. The Starter is model-agnostic in concept, but the examples assume Claude + Cowork.
How technical do I need to be? ›
Comfortable editing markdown files and running occasional CLI commands. You don't need to write code. If you can manage a Notion workspace, you can run this.
How long does setup take? ›
30 minutes for the Starter shell. 1-2 weeks of honest use before the benefits compound. Most people feel a difference inside week one.
What's the difference between the Starter and the 1:1 build-out? ›
Starter = skeleton. Build-out = memory tuned to your work, router for your projects, scheduled tasks for your rhythms, observability for your edge cases, and the rules re-derived from your actual failure modes.
Is my data safe if I share it with Claude? ›
Claude's enterprise and Pro plans don't train on your data. The Starter keeps all memory local in your files. Treat memory files like you'd treat a private git repo.