Summary
Nate Herk unpacks insights from an interview with Boris Churnney, one of Claude Code’s creators at Anthropic, revealing how the team approaches system prompt management across model generations. The central finding: Anthropic deletes over 80% of Claude Code’s system prompt with each new model release, treating each model as a distinct entity with different strengths, flaws, and prompting needs. For Opus 5 specifically, Anthropic recommends stripping away accumulated CLAUDE.md files, skills, and hooks to see what the model can handle natively — a counterintuitive move for power users who have spent months refining their setups.
Herk validates this live by duplicating his workspace and running parallel tests with and without his established skill files. His finding: the unconstrained model often produces conceptually richer outputs — in one example generating a better-structured resource guide with timestamps — even if it lacks personal formatting preferences baked in over time. His recommendation is surgical rather than wholesale: preserve style and brand-specific rules, but strip instructional scaffolding the newer model no longer needs.
He closes with an important caveat about advice provenance: guidance from engineers like Churnney or Andrej Karpathy comes from people building large-scale software harnesses, not doing knowledge work or content creation. Viewers should calibrate AI workflow advice against how they actually use these tools day-to-day rather than treating it as universal.
📺 Source: Nate Herk | AI Automation · Published August 12, 2026
🏷️ Format: Opinion Editorial
