A Juejin author writing as "Silent Wang Er" walks through Claude Code's internals based on its official documentation, starting with the seven kinds of content in the context window and the two-step overflow strategy: clearing older tool outputs first, then summarizing the conversation history, which the author notes can lose earlier instructions. The piece details how four CLAUDE.md layers (managed policy, user, project, local) are concatenated rather than overridden, with an official recommendation to keep each file under 200 lines, and how Hooks enforce rules via exit codes and PreToolUse permission decisions such as allow, deny and ask. It also covers Skills' on-demand loading, a 1536-character cap on each Skill description, a 1% context budget for the skill index list, and the context isolation of the built-in Explore, Plan and General-purpose Subagents, where Explore and Plan skip loading CLAUDE.md and Git status at startup.