Claude Code의 진짜 힘은 CLAUDE.md가 아니라 스킬 시스템에서 나옵니다.
필요한 순간에만 스킬 파일을 불러오는 점진적 노출 방식이, 모든 컨텍스트를 항상 로드되는 단일 파일에 몰아넣는 방식을 이깁니다.
Claude Code's real power isn't CLAUDE.md — it's the skills system. Progressive disclosure (loading skill files on demand) wins over front-loading every rule into one always-loaded context file.
ai-codingmemory-systemssv-developer
Claude Code의 진짜 힘은 CLAUDE.md가 아니라 스킬 시스템에서 나옵니다.
필요한 순간에만 스킬 파일을 불러오는 점진적 노출 방식이, 모든 컨텍스트를 항상 로드되는 단일 파일에 몰아넣는 방식을 이깁니다.
CLAUDE.md 담론의 한가운데에서 실밸개발자가 던지는 반전 주장입니다. 근거는 단순합니다 — 컨텍스트 윈도우가 한정되어 있기 때문입니다. CLAUDE.md는 매 세션 100% 로드되지만, 스킬은 호출될 때만 로드됩니다. 그래서 같은 컨텍스트 예산으로 더 많은 전문 지식을 굴릴 수 있는 구조는 스킬입니다. Karpathy의 "스키마가 가장 큰 레버리지"를 정정합니다: 스키마는 필요 조건이고, 점진적 노출 시스템(스킬)이 실제 레버리지입니다.Also in EnglishClaude Code's real power isn't CLAUDE.md — it's the skills system. Progressive disclosure (loading skill files on demand) wins over front-loading every rule into one always-loaded context file.
A counter-positioning move at the center of the CLAUDE.md discourse. The reason is simple: the context window is finite. CLAUDE.md loads 100% every session; skills load only when invoked. So the structure that fits the most expertise into the same context budget is skills, not CLAUDE.md. A refinement of Karpathy's "schema is the highest-leverage artifact" — the schema is necessary, but the progressive-disclosure system (skills) is the actual leverage.