One-Person System (OPS)

Karpathy sketched the LLM as an OS. OPS is the personal filesystem that runs on top of it — and survives every model swap.

Jarvis is coming. Will he actually know you? An open spec for the structured "you" that no vendor owns.

中文版 README →

The Question OPS Answers

Personal AI assistants are coming. Every cloud vendor is building one. Within 3 years, "everyone has a Jarvis" will be a reasonable statement.

But:

Jarvis is coming. Will he actually know you?

OPS is the answer. It's the framework that lets one human individual structure their principles, identity, cognition, memory, and outputs such that any sufficiently capable AI — current or future, vendor or self-hosted — can read them, query them, and assist meaningfully.

Without becoming the author of your judgment.

What OPS Is

OPS — One-Person System — is the first formal proposal of an architecture that:

Treats the individual as the unit of system design (not the company, not the team)
Is AI-native from day one (designed to be read, queried, compiled, and assisted by AI agents)
Enforces data sovereignty before model sovereignty — your structured self survives any vendor change (P6 Model Portability)
Treats privacy as architectural primitive — every data unit has a room label (Bedroom / Living Room / Outside) (P7)
Maintains Pilot Sovereignty — the human owns judgment, AI is interchangeable interface (Pilot Doctrine)

OPS is not a tool. It is a specification. Any concrete implementation (Obsidian, Notion, plain text, custom) that meets the conformance criteria is a valid OPS.

📜 Read the spec: SPEC.md (English) · SPEC.zh.md (中文) 🔥 Read the manifesto: MANIFESTO.md (English) · MANIFESTO.zh.md (中文)

Beyond Karpathy's LLM OS

Karpathy's LLM OS sketch put the model at the center — LLM as kernel, tools as peripherals, context as RAM. It's a system architecture for the machine side.

OPS is the human-side counterpart — the structured user that the LLM OS operates on.

Karpathy's LLM OS	What OPS adds
LLM = kernel, tools = peripherals	A user-side spec — L0–L5 layered filesystem for the human
System architecture sketch	A governance contract — the Pilot Doctrine (8 rules), human owns judgment
Privacy not addressed	3-Room Privacy as architectural primitive (P7) — every datum has a room
Model = center of intelligence	Data + algorithms = persistent self; the model is interchangeable (P6)
Forward-looking research vision	Runnable today on Obsidian + Claude Code — 25–35 min to MV-OPS

One line: Karpathy's LLM OS is the kernel. OPS is the home directory.

Architecture at a Glance

┌─────────────────────────────────────┐
│  L5  SYSTEM       compiler / index  │ ← AI lives here
├─────────────────────────────────────┤
│  L4  EXECUTION    modules / output  │ ← Your outward work
├─────────────────────────────────────┤
│  L3  DATA         raw + wiki        │ ← Your long-term memory (with rooms)
├─────────────────────────────────────┤
│  L2  COGNITION    methodologies     │ ← Your judgment engine
├─────────────────────────────────────┤
│  L1  IDENTITY     dynamic portrait  │ ← Your self-knowledge
├─────────────────────────────────────┤
│  L0  CONSTITUTION principles        │ ← Pilot-only · immutable
└─────────────────────────────────────┘

Three rooms (privacy tiers) on every data unit:

🚪🚪🚪 Bedroom (卧室)      local AI only
🚪🚪   Living Room (客厅)  cloud AI permitted, no training/cache
🚪    Outside (门外)       public, no restriction

Four loops of operation: Ingest → Compile → Query → Reflect.

Eight Pilot Doctrine rules. The first and the last:

Rule 1: AI is not your co-pilot. You are AI's pilot. Rule 8: Your data, borrowed models. Bedroom is yours. Model is rented. Never confuse the two.

🔑 The core claim — your data outlives the model. Switch from Claude to ChatGPT, Kimi, GLM / Zhipu, DeepSeek, Qwen, or a local Llama tomorrow. Your structured "you" stays. Intelligence lives in your data and algorithms — not in any vendor's weights.

Two Editions (Phase 1)

OPS comes in two packaged editions, by user experience density:

🌱 OPS-Starter-Edition

For: students, early-career professionals, side-hustle starters, AI/PKM newcomers, transition-stage individuals

2-room privacy (Bedroom / Outside)
5-7 question L0
1 entity type to start
90-day to MV-OPS
Teaching-toned UI

📁 editions/starter-edition/

🎯 OPS-Executive-Edition

For: decision-makers — C-level, founders, partners, government leaders, family-business heirs

3-room privacy (Bedroom / Living Room / Outside)
12-question L0 (includes power-boundary + vested-interest questions)
4 entity types from start (people / orgs / decisions / events)
30-day to MV-OPS
Sharp, professional UI

📁 editions/executive-edition/

Long-term roadmap will split into 7 editions (Operator / Thinker / Heir / Seeker / Doer / Explorer / Capital) — see docs/personas.md.

See It Working in 60 Seconds

Before you read the spec — see a working OPS.

git clone https://github.com/Leohuangcn/leo-ops.git

Open leo-ops/starter-vault/ in Obsidian as a vault. You're now looking at a complete, runnable OPS skeleton — all six layers (L0 Constitution → L5 System), all entity templates, all rooms wired in. Click into 01-本体/本体-模板.md or 99-系统层/CLAUDE.md to see how the pieces talk to each other.

Then decide whether to read the spec.

Five Ways to Engage

1. Read — The Specification

If you want to understand OPS:

SPEC.md / SPEC.zh.md — full v0.2 specification (~3500 words)
MANIFESTO.md / MANIFESTO.zh.md — why this exists, in 2 minutes
docs/method-deconstructing-self.md — the 3-dimensional self-labeling method

2. Build — The Starter Vault

If you want to start your own OPS today:

git clone https://github.com/Leohuangcn/leo-ops.git
cp -r leo-ops/starter-vault ~/my-ops
# open ~/my-ops in your editor of choice

3. Run — The Claude Skill (recommended guided co-pilot)

If you use Claude Code:

cp -r leo-ops/skills/ops-builder ~/.claude/skills/
# in Claude Code:
/ops-builder
# language selection → guided onboarding in 25-35 minutes → MV-OPS

4. Run — Any Other AI (ChatGPT / Cursor / Kimi / GLM / DeepSeek / Qwen / local Llama)

OPS Spec v0.2 P6 (Model Portability) means OPS works with any sufficiently capable AI.

For non-Claude runtimes:

Open your OPS vault folder in the AI tool if it supports files/projects.
Give the AI 99-系统层/CLAUDE.md as its operating rules.
Ask it to follow the four loops: Ingest → Compile → Query → Reflect.
Keep Bedroom data local; only Living Room / Outside data may enter cloud model context.

Chinese cloud models such as Kimi, GLM / Zhipu, DeepSeek, and Qwen can operate OPS as long as they follow the same vault rules and room boundaries.

5. Pick Your Edition

Don't use the generic starter — pick the edition matching your stage:

Starter Edition for low experience density
Executive Edition for decision-makers
See docs/personas.md for full persona-to-edition mapping

What's New in v0.2

	v0.1	v0.2
Principles	P1-P5	P1-P9 (added Model Portability, Three-Room Privacy, Pilot-only Labeling, Decision Reflexivity)
Pilot Doctrine	7 rules	8 rules (added "Your data, borrowed models")
Anti-patterns	AP-1 to AP-5	AP-1 to AP-9
Privacy	not specified	3 rooms architecture mandatory
Self-labeling	not specified	3-dim ontology specified (see method doc)
Editions	none	2 editions (Starter / Executive)
Languages	bilingual docs	fully bilingual (spec + skill + templates)
Skill onboarding	scaffold	complete 5-stage flow (25-35 min to MV-OPS)

Full changelog: CHANGELOG.md

Prior Art & Differentiation

OPS does not invalidate prior work. It builds on:

Prior Work	What OPS Adds
PARA (Tiago Forte)	A principles layer, AI-native design, layered isolation
Life OS (August Bradley)	A formal specification, tool independence, Pilot contract
Building a Second Brain	The Pilot Doctrine — human authority over machine assistance
One-Person Company (OPC)	The underlying individual operating system that OPC is one module of
Karpathy's LLM-OS / Personal AI vision	Data sovereignty before model sovereignty — OPS stores intelligence in data + algorithms, not model weights

Full comparison: docs/comparison.md · OPS Spec § 1.4

Status

v0.2.0 — Public Draft. This specification is open for public review. Feedback, issues, and PRs welcome.

Particular interest in:

Conformance edge cases
Non-Obsidian implementations (Notion, Logseq, Roam, custom)
Alternative AI runtimes (Cursor, ChatGPT Projects, custom agents, local model stacks)
Translations beyond zh / en

Roadmap

Version	Focus	Target
v0.1 ✅	Initial proposal, starter vault, Claude skill scaffold	2026-05-25
v0.2 ✅	Three rooms, model portability, decision reflexivity, bilingual, complete onboarding skill, 2 editions	2026-05-25
v0.2.1 ✅	Multi-AI portability surfaced in product docs; L5 indexes + deterministic scripts; privacy/docs/CI hardening	2026-06-19
v0.2.2	Runtime-specific prompt packs (ChatGPT / Cursor / Kimi / GLM / DeepSeek / Qwen)	Q3 2026
v0.3	Pilot directory, certification mechanism, additional edition splits, community translations	Q4 2026
v1.0	Stable specification, formal governance, multi-stakeholder maintenance	2027

Citation

If you cite OPS in writing or research:

Huang, Leo (黄乐). One-Person System (OPS) Specification, Version 0.2.0.
2026-05-25. https://github.com/Leohuangcn/leo-ops

Machine-readable: CITATION.cff

License

Specification & documentation (SPEC*.md, MANIFESTO*.md, spec/, docs/, editions/*/README.md): CC-BY-4.0 — share and adapt with attribution
Code, skills, vault templates (skills/, starter-vault/, scripts): MIT

Author & Contact

Leo Huang (黄乐) · @Leohuangcn

Issues / bugs: GitHub Issues
Discussion / ideas: GitHub Discussions
WeChat newsletter (Chinese): 「Leo 的 AI 数字文化笔记」

A Final Word

Models will turn over. Compute will become commodity. Everyone will have their Jarvis.

But will Jarvis know you?

You have ~3 years to make sure the answer is yes — by building the structured "you" that no vendor owns.

That's what OPS is for.