Reflection: If You Could Redesign It, What Would You Keep

Wisdom lives in the prompt, not in the code: How Claude splits tasks and which tools it uses are taught to the model via the system prompt — the code only handles execution and routing. This means behavior can be iterated without changing code.

Context compression is the foundation of commercial viability: Without compression, Claude Code would constantly overflow its context on large projects and couldn't complete complex tasks. The three-layer compression strategy is the fundamental guarantee that this product actually works.

Worktree isolation = reduced risk from AI operations: Having agents work inside git branch copies is an elegant safety design — it contains "AI might break the code" risk at the git level.

Feature-flag-driven product evolution: Large numbers of features (KAIROS, PROACTIVE, COORDINATOR_MODE, etc.) are gated by feature flags, showing Anthropic continuously runs A/B tests and progressively opens new capabilities to users.