Liu's 4 Lines Are the Floor. Build the Ceiling.

The core concept of effective agent configuration lies in prioritizing behavioral rules over specific features. This principle, articulated by Yanli Liu and rooted in Andrej Karpathy's work, emphasizes that agents should focus on broad principles rather than rigid rules. The suggested "4 lines" offer a strong starting point for CLAUDE.md files. However, the author argues that these lines are insufficient on their own without architectural considerations, that a production CLAUDE.md should be an architecture, not a single file. The author highlights the "Configuration Paradox," where excessive rules can lead to agent confusion and ineffectiveness. This also promotes the use of delegated files for project-specific rules, domain context, and incident-driven corrections. The proposed architecture involves separating behavioral guardrails, project-specific contexts, language-specific rules, and lessons learned with specific file types. This approach promotes modularity and reduces complexity, making the agent more efficient. The author suggests filters for determining rule placement, based on project, language, or incident specificity. This design minimizes the agent's working set by loading only relevant information. The architecture addresses limitations in multi-file refactoring, domain-specific guardrails, and team-scale consistency. The author emphasizes that the "4 lines" offer a solid foundation but the overall architecture is much more important. The ultimate goal is to create CLAUDE.md files that serve as dispatchers, not comprehensive rulebooks, improving agent performance.