The original title is: "Auditing a production Claude Code setup against Anthropic's context-engineering rules"

Short summary

A team audited their Claude Code setup against Anthropic's published context-engineering rules and found their session-start hooks were injecting the same 60-line preferences list three times per session via overlapping gps commands. The audit revealed violations of lightweight CLAUDE.md, progressive disclosure, and automatic memory principles. The post details the root cause analysis and fixes applied to reduce context bloat before any real work begins.

• •Session-start hooks were injecting the same 60-line preferences list three times per session via redundant gps commands
• •105 preference entries included out-of-context code comments and test fragments rather than genuine standing invariants
• •Audit maps directly to Anthropic's five rules: keep CLAUDE.md lightweight, progressive disclosure, trust model judgment, rely on automatic memory, design tools not prose

Generated with AI, which can make mistakes.