Summary
Cole Medin unpacks a viral moment from a Y Combinator event where Boris Churnney, creator of Claude Code, told developers to delete their entire AI layer — CLAUDE.md files, skills, hooks — every six months and see what modern models can handle without instruction. Medin argues this advice has been widely misread as a mandate to trust the model entirely, when Churnney was actually describing a disciplined research practice called ablation: remove everything, then add back only what demonstrably improves output.
Medin structures the video around three components of an AI layer — global rules, skills and workflows, and sub-agents — and argues only the rules layer deserves aggressive pruning. He cites Anthropic’s own Claude Code documentation, which recommends keeping CLAUDE.md files under 200 lines and explicitly warns that over-specified global rules cause important instructions to get lost in the noise. Skills and sub-agents, by contrast, solve real coordination problems that even Opus 5 cannot self-organize around.
The video also introduces two PR review tools — Code Rabbit (free for open source) and ChangeStack (free during launch) — demonstrated on Medin’s Archon project, where Code Rabbit identifies real issues and supports conversational clarification. The overall message is practical: yes, trim bloated global rules as models improve, but wholesale deletion is expensive, slow, and often unnecessary. Ablation is a diagnostic tool, not a deployment strategy.
📺 Source: Cole Medin · Published August 06, 2026
🏷️ Format: Opinion Editorial
