AI Architecture
3 posts
When AI's "Brain" Isn't Enough——Compaction and Microcompact
Deep dive into Claude Code’s context compaction mechanism: differences between auto-compaction and microcompact, state preservation after compaction, and context collapse. Unveiling how AI elegantly makes tradeoffs when “memory runs out,” and why AI sometimes “forgets” previous content.
System Prompts——The "Director Behind the Scenes"
Deep dive into Claude Code’s system prompt architecture: segmented composition design, division of identity/tools/format/safety modules, and the secrets of model-specific tuning. Unveiling why the same request gets completely different response styles from Claude Code versus ChatGPT.
Agent Loop——How Does an AI's "Brain" Work?
Deep dive into Claude Code’s core Agent Loop: the complete lifecycle from user input to model response. Unveiling the query.ts state machine, tool execution orchestration, and continuation decision logic to understand how the AI coding assistant “thinks” and “acts.”
