Your organization already has an operating system for AI. You just haven't written it down.
It exists in the CLAUDE.md file you maintain. In the Slack messages where you corrected an agent's behavior. In the mental model you carry about which agent handles what, and what happens when two of them disagree. In the rules you discovered the hard way after an agent sent the wrong email or overwrote the wrong data.
That implicit operating system is the most valuable artifact in your AI stack. More valuable than the model. More valuable than the prompts. Because it encodes the coordination intelligence that took you months to build.
The OOS makes it explicit.
From CLAUDE.md to OOS
If you use Claude Code, you probably have a CLAUDE.md file. It's the set of instructions that tells Claude how to behave in your project. Roles, rules, tool access, escalation paths, things not to do.
A CLAUDE.md is a good start. But it has limitations:
- No confidence ratings. Every rule looks equally certain, whether it's a hard-won lesson from a production incident or a guess you added last Tuesday.
- No evidence typing. Did you measure this? Observe it once? Infer it from theory? The CLAUDE.md doesn't say.
- No failure modes. What goes wrong when this rule is violated? Without documented failure modes, new team members (human or AI) don't understand the stakes.
- Not portable. You can't diff two CLAUDE.md files and get structured insights. You can't merge them. You can't search across them.
- Not versioned meaningfully. Git tracks changes, but the evolution of your coordination intelligence isn't captured in a way that shows growth, regression, or stagnation.
The OOS takes everything in your CLAUDE.md and structures it as claims. Each claim is a discrete unit of coordination intelligence.
Anatomy of a Claim
A claim is the atomic unit of organizational intelligence. Here's what one looks like:
Every agent writes to its own shared state file. No agent modifies another agent's state.
Why: Prevents data corruption from concurrent writes.
Failure mode: Two agents write to the same file. One update is lost. State becomes inconsistent.
Scope: All agents. Each agent has exactly one state file it owns.
Six fields. Each one does specific work:
Rule: What the organization actually does. Not aspiration. Operational reality.
Why: The reasoning. This is what makes intelligence transferable. Without the "why," a rule is just a policy. With it, another organization can evaluate whether the reasoning applies to them.
Failure mode: What happens when this rule breaks. The scar tissue. This is the most underrated field. It turns abstract rules into visceral consequences.
Confidence: HIGH, MEDIUM, or LOW. An honest assessment of how certain the organization is. HIGH means battle-tested. LOW means "we think this is right but haven't proven it."
Evidence: How the organization knows. MEASURED_RESULT (data), OBSERVED_REPEATEDLY (pattern), OBSERVED_ONCE (anecdote), HUMAN_DEFINED_RULE (policy), INFERENCE (logical), SPECULATION (guess).
Scope: Where this applies. All agents? Just the email agent? Only during onboarding?
What Sections Look Like
An OOS organizes claims into sections. The Agent Army template (the most common) uses these:
Why Structure Matters
You might think: "I could just share my CLAUDE.md." You could. But structured intelligence does things unstructured text cannot:
You can diff it. Put two OOS files side by side and the comparison engine shows: what's unique to each, what overlaps, and where they conflict. Try doing that with two CLAUDE.md files -- you get a wall of text.
You can search it. "Show me every claim about shared state architecture with HIGH confidence and MEASURED_RESULT evidence." That query works across every published OOS. You can't do that with raw markdown.
You can merge it. Import claims from another organization at LOW confidence. Review them. Adjust confidence as you validate in your own environment. Your OOS grows with borrowed intelligence.
You can track its evolution. Version 1 had 12 claims, 4 HIGH confidence. Version 3 has 24 claims, 10 HIGH. Your coordination intelligence is measurably growing.
How to Generate Yours
If you have a CLAUDE.md or similar agent configuration, you can generate an OOS in about 30 minutes. The process:
- Extract claims from your existing config. Every rule, every boundary, every pattern becomes a claim. Claude can help -- paste your CLAUDE.md and ask it to extract claims in the OOS format.
- Rate each claim honestly. Don't mark everything HIGH confidence. If you haven't tested it, it's LOW. If you observed it once, it's MEDIUM at best. Honesty makes the dataset valuable.
- Document failure modes. For each claim, answer: "What happens when this breaks?" If you can't answer, the claim might be theoretical rather than operational.
- Publish. Paste your OOS into the publisher on OTP. The platform validates format, scans for PII, calculates your quality tier, and publishes you to the Intelligence Graph.
Step-by-step instructions with examples
Create your OOS and join the Intelligence Graph
The Bigger Picture
One OOS is a snapshot of how one organization coordinates its AI. Useful for that organization, but limited.
Many OOS files, published to the same graph, create something none of them could create alone: a map of how humans and AI agents actually work together across industries, team sizes, and use cases.
That map doesn't exist today. Every organization is discovering coordination patterns independently. The same shared-state architecture. The same escalation protocols. The same human-AI boundaries around financial decisions and client communication.
The OOS is how we build that map. One claim at a time.
Founder of OTP and CEO of Sneeze It, a digital marketing agency running 14 AI agents in production.
dsteel@sneeze.it