Type: url
Original: https://zenn.dev/genda_jp/articles/f71d3ed7d4d7e8
Status: generated
Summary
This post argues that the instruction files handed to AI coding agents have split into three layers with non-overlapping concerns: AGENTS.md/CLAUDE.md for an agent’s overall premises and boundaries, SKILL.md for reusable individual tasks, and DESIGN.md — published by Google Labs in April 2026 with an accompanying lint CLI — for design-system specifications. It reads the split as a question of where each layer places the balance between machine-readable and human-readable content, and offers it as a criterion for deciding how to break up a single overloaded CLAUDE.md. It then sets this against spec-driven development, finding a shared philosophy but a different time axis: SDD specs describe what is about to be built and are archived once done, while the three-layer files describe standing norms that are maintained and grow.