Summary
Anthropic engineers recently published new guidance on context engineering for Claude 5-generation models, and the team reports cutting more than 80% of Claude Code’s system prompt with no measurable loss on their evaluations. This video from Nate Herk walks through what that change means for anyone who has built up long instruction files, skills, and CLAUDE.md rules over time.
The core argument is that rules written to compensate for older models can now hold newer ones back. Herk explains Anthropic’s advice to use the smallest amount of context that still gets the result you want, while noting that minimal does not mean short: details the model genuinely needs should stay. He cites research showing performance can drop as input length grows, and Anthropic’s own results from loading MCP tool definitions only when needed, which cut token usage by 85% and lifted evaluation scores for Opus 4 and Opus 4.5.
The second half is a practical demo. Herk uses Claude to audit his own content-creation workflow, runs the /doctor command to surface duplicated or outdated instructions, and rewrites a heavily constrained skill so the model can use its own judgment. Viewers come away with a checklist for pruning skills, deciding what context to keep, and defining what done looks like instead of dictating every step.
📺 Source: Nate Herk | AI Automation · Published October 08, 2026
🏷️ Format: Tutorial Demo
