정리 1
Context is King
단, "다 던지기"는 함정 — Fresh + Condensed 큐레이션이 답
정리 2
Probabilistic Confusion
LLM의 확률 엔진 특성에 노이즈가 더해질 때 variance가 폭발
정리 3
Second Brain
specs · decisions · rationales의 hierarchical markdown tree, CLAUDE.md 진입점
정리 4
OpenAI Harness 원칙
"코드베이스에 없으면 agent에겐 존재하지 않는다" → tacit를 codify
정리 5
Toolset 3종
Slash Commands · Sub-agents/Skills · MCP — 통합 시스템으로
다음 · 클립 03 · Pillar II
Agentic Validation — agent의 "done"을 누가 검증하는가
Context를 정밀하게 큐레이션해도, agent의 결과는 독립적으로 검증되어야 한다 — 그 검증 시스템을 다음 클립에서