Governance Templates for Claude Cowork

Practical governance templates for long-horizon work with Claude Cowork.

Cowork gives Claude filesystem access. These templates give you the structure to make that sustainable.

---

The Problem

Cowork is powerful. But power without governance becomes chaos.

The AI doesn't "remember" — it reconstructs. Every session, it rebuilds context from whatever's in front of it. Without structure:

• Files start quietly contradicting each other
• Numbers get reconstructed instead of referenced
• Decisions fragment across versions
• Drift is invisible until it isn't

Week one feels like magic. Week three, you're untangling a mess.

---

The Solution

These templates implement a minimal governance layer built from a year of daily human-AI collaboration. They address the failure modes that emerge in long-horizon work — not by preventing failure, but by making it visible, bounded, and repairable.

---

What's Included

Template	Purpose
RUNNING-DOCUMENT.md	Session continuity — read at every session start
CANONICAL-NUMBERS.md	Single source for numeric truth
FOLDER-STRUCTURE.md	How to organize your Cowork folder
FAILURE-RECOVERY.md	What to do when things break

---

Quick Start

1. Create your Cowork folder
2. Copy these templates into it
3. Read FOLDER-STRUCTURE.md and organize accordingly
4. Start every session by pointing Claude to your RUNNING-DOCUMENT.md
5. Keep all numbers in CANONICAL-NUMBERS.md — never let Claude reconstruct them

That's it. Boring. Unglamorous. Effective.

---

The Research Behind This

These templates are derived from a four-paper research series on governed human-AI collaboration:

• Paper 1: Context-Engineered Human–AI Collaboration for Long-Horizon Tasks
• Paper 2: The Lean Collaboration Operating System (LC-OS)
• Paper 3: Failure and Repair in Long-Horizon Human–AI Collaboration
• Paper 4: The Living Framework: Living with a Governed Human-AI Dyad

All open access: github.com/LivingFramework/LC-OS

---

Core Principles

1. Reliability isn't preventing failure — it's structured repair
2. Single Source of Truth per domain — no parallel versions
3. Separate text from numbers — they drift differently
4. Read the Running Document at every session start
5. When in doubt: Stop → Diagnose → Rollback → Note

---

Author

Rishi Sood Independent researcher focused on human-AI collaboration governance.

• Research: github.com/LivingFramework/LC-OS
• X: @satoshisats

---

License

MIT — use freely, attribution appreciated.