Descriptions:
Nate Herk walks through five techniques for keeping an AI Operating System (AI OS) accurate and scalable as the knowledge base grows, responding to widespread audience questions about organizing Claude Code agent projects. The video opens with a live demonstration of a free downloadable ‘OS Audit’ skill that scans an entire Claude project, checks every routing rule and index claim against the actual file structure, and produces a prioritized fix report without making any modifications.
Before covering the five tips, Herk introduces a taxonomy of four context failure modes: poisoning (a false fact embedded in context), bloat (too much data for the model to navigate effectively), confusion (contradictory instructions), and clash (conflicting routing rules). He pairs this with a distinction between ‘expertise context’ — stable background knowledge loaded at all times — and ‘situational context’ loaded on demand for specific tasks. Understanding these failure types, he argues, is prerequisite to diagnosing why agents produce incorrect outputs at scale.
The practical tips center on using CLAUDE.md as a router rather than a monolithic system prompt, structuring wikis and indexes so they mirror the actual state of the project on disk, and deploying sub-agents to parallelize large audit tasks across projects with more than 100 folders. The audit skill checks routing integrity, index truth, data freshness, memory bloat, and duplication in sequence, and is available for free inside Herk’s School community. The techniques are designed to work with any Claude Code project regardless of domain.
📺 Source: Nate Herk | AI Automation · Published July 23, 2026
🏷️ Format: Tutorial Demo
