Everyone can bolt a chatbot onto a product. Almost no one architects a company where intelligence is the core, where every interaction makes the system smarter and the thing falls apart without it. AI-Native OS is the whole method, open and runnable: a 15-chapter Handbook, a 90-entry Dictionary, and 25 AI skills that take you from a vague idea to a company with a real data moat.
Handbook · Dictionary · Skills · Install · Non-coders start here · Contribute
Part of a family. AI-Native OS has a sibling for the general founder journey: Founder OS. Same author, same licence. Use Founder OS for idea-to-operations; use AI-Native OS when your product itself is the AI.
For 25 years I worked in food and biotech R&D, the kind of science where a confident guess that skips the evidence costs millions. When I started building companies, I watched a new mistake replace the old one: founders shipping an LLM call wrapped in a UI and calling it an "AI startup." It demos beautifully. Then a competitor ships the same wrapper in a weekend, because there was never a moat, just an API key.
The companies that last are built the other way around. Intelligence sits at the core: the product is a closed loop that gets smarter with every use, the org runs on agents instead of headcount, and the data it accrues can't be copied. That is an AI-native company, and almost nobody teaches how to build one, least of all in the hard-mode sectors (food, health, deeptech) where the regulation is real and the demos can kill someone.
So I built the operating system I wish I'd had. AI-Native OS turns "I want to build with AI" into a sequence of clear, architecture-first steps: frame the bet, design the cognitive system before you touch code, build with agents, earn trust in sensitive domains, ship a Minimum Viable Agent, grow via GEO, and defend the moat. It is theme-agnostic on purpose (you supply the domain, it supplies the method) and open, because the AI-native playbook should not be a $5k course.
- Adam M. Adamek, PhD · Impact Brussels ASBL · building in the open Read the MANIFESTO · meet the author and why I write.
The one test this whole repo is built around, the Remove-the-AI test: take the AI out of your product. If it still works, you built an AI feature, not an AI-native company. Everything here pushes toward systems that break without their intelligence, because that is where the moat is.
AI-Native OS is organised the way an AI-native company actually gets built:
Frame → Architect → Build → Trust → Ship → Grow → Defend
Not sure where you are? Run start-here: it asks a few questions, runs the Remove-the-AI test,
and routes you to the one skill to run next. Full map: docs/STAGES.md.
AI-Native OS has three layers that reinforce each other:
| Layer | What it is | Where |
|---|---|---|
| Handbook | Load-Bearing, the 15-chapter method (00 plus 14), long-form, ~33,800 words, with named case studies (NotCo, Foodpairing, OpenEvidence, Nuritas, BeeWise, Impact Brussels…). | handbook/ |
| Dictionary | The AI-native vocabulary: 90 terms in 7 categories, each defined the way Matt Pocock defines them: precise, opinionated, with usage and what to avoid. | dictionary/ |
| Skills | 25 skills, one founder job each, every one a one-stop shop with its own references/ folder. Runnable on Claude Code / Codex / Cursor / Gemini, or copy-pasted into any chatbot. |
skills/ |
Plus 5 flows (multi-step workflows with checkpoints), 5 scheduled loops with a
loop-engineering subsystem (docs/LOOP-ENGINEERING.md and the
design-a-loop skill), 6 advisor agents (a structured devil's
advocate, a safety judge, a clinical reviewer, a regulatory proxy, and more), 26 copy-paste
prompts (prompts/), masterfiles (drop-in system context), and a living
knowledge base.
The method, in depth: 15 chapters (00 plus 14), ~33,800 words. The handbook is titled "Load-Bearing" (full title: Load-Bearing: Building AI-Native Companies in Hard-Mode Sectors); the repo and OS itself stay AI-Native OS. Each chapter opens with the problem, gives you the core concepts, shows 2–3 real companies that did it, and ends with the exact AI workflows (which tool, which prompt) plus a "Your turn" exercise. Read it start to finish, or jump to your stage.
| # | Chapter | Stage |
|---|---|---|
| 00. | What "AI-native" actually means | (all) |
| 01. | The AI-native founder in hard-mode sectors | Frame |
| 02. | From vague idea to testable hypothesis | Frame |
| 03. | Mapping markets, competitors & EU terrain | Frame |
| 04. | Customer discovery that doesn't lie to you | Frame |
| 05. | Architecting your system before you touch code | Architect |
| 06. | Agentic coding in practice (Claude Code, Codex, Cursor, Gemini) | Build |
| 07. | Managing agentic technical debt before it kills you | Build |
| 08. | Evaluation & safety for multi-agent systems | Trust |
| 09. | Designing AI-native MVP experiments, not features | Ship |
| 10. | Narrative, positioning & founder brand (GEO) | Grow |
| 11. | AI-assisted sales, pilots & enterprise motions | Grow |
| 12. | Measuring real product-market fit, not AI hype | Grow |
| 13. | AI-native ops: automating the boring stuff | Defend |
| 14. | Moats, data & ecosystems | Defend |
You can't build what you can't name. The Dictionary defines the 90 words an AI-native
founder lives in (context window, agent, MCP, hallucination, CACE, GEO, Share of Model, Minimum Viable Agent, human-on-the-loop, data flywheel), across 7 categories, each with
a one-line definition, the mechanics, what to avoid, and a real usage example. The Handbook and the
skills link into it, so the vocabulary is one coherent graph rather than scattered jargon.
All 25 skills are shipped. Each is one job, produces a real artefact, carries its own references/
folder so it is a one-stop shop, and ends with a copy-paste prompt for non-coders.
| Stage | Skills |
|---|---|
| Frame | frame-the-hypothesis · map-the-terrain · customer-discovery-that-doesnt-lie · hypothesis-mining |
| Architect | architect-before-code · cognitive-architecture-review |
| Build | agentic-build-loop · pay-down-agentic-debt · write-the-claude-md · self-healing-fallbacks |
| Trust | eval-and-safety-harness · red-team-the-agent · secure-the-connectors · compliance-readiness |
| Ship | design-the-mva |
| Grow | measure-ai-native-pmf · geo-content · share-of-model-audit · ai-assisted-sales |
| Defend | gateway-agent-ops · moat-strategy |
| Cross-cutting | start-here · capture-learning · apply-ai-native-models · design-a-loop |
Plus 5 flows (idea-to-mva-flow, ship-with-confidence-flow, architecture-first-flow,
geo-launch-flow, fundraise-the-ai-native-way-flow), 5 scheduled loops (daily-agent-standup,
ci-failure-triage, weekly-share-of-model-review, weekly-debt-and-eval-review, red-team-cadence)
with a loop-engineering subsystem, and 6 advisor agents (devils-advocate, safety-judge,
clinical-reviewer, regulatory-proxy, data-flywheel-architect, customer-proxy). What's next is
in ROADMAP.md: a print-ready Handbook build, a worked sample company, more loops, and
translations, each a ready contribution slot.
Fastest (any agent), one command, no clone, works across Claude Code, Cursor, Codex and more:
npx skills add impactbrussels/AINativeOSClaude Code, or clone into your skills path (or add as a plugin):
git clone https://github.com/impactbrussels/AINativeOS.git
cd AINativeOS
./install.shThen, in Claude Code: "Use start-here. I want to build an AI-native company." The installer wires the skills into your tool; the full walkthrough, per platform, is in INSTALL.md.
Codex / Cursor / Gemini: the same skills/ source works. Codex reads AGENTS.md,
Cursor reads .cursor/rules/ai-native-os.mdc, Gemini reads
GEMINI.md. See the cross-platform guide.
You do not need any developer tools. Every skill has a copy-paste prompt at the bottom, and the
Prompt Library collects all 26 as standalone files. Pick your stage, copy the
block, paste it into Claude.ai / ChatGPT / Gemini (or into a no-code builder like Lovable, Bolt,
v0, or Replit), fill the [PLACEHOLDERS], and iterate. New to all of this? Start with the no-code
guide: docs/FOR-NON-TECHNICAL-FOUNDERS.md.
AI-Native OS is built to be built with you. The roadmap lists the planned skills,
flows, loops, and dictionary entries. Copy the gold-standard architect-before-code
skill, follow CONTRIBUTING.md, and open a PR.
Dual-licensed so credit is legally required:
- Content (handbook, dictionary, skills, prompts, docs): CC-BY-4.0.
- Code (scripts, workflows): Apache-2.0.
Attribution: AI-Native OS by Adam M. Adamek (Impact Brussels ASBL). Details: ATTRIBUTION.md. Citing in research? See CITATION.cff.
Credits and related work: the thinkers, tools, and prior art this OS builds on are acknowledged in CREDITS.md.
AI-Native OS · made for founders, by a founder · Impact Brussels ASBL