An interview with Boris Cherny, the creator of Claude Code, with a high density of valuable insights. A few memorable points: Claude Code uses grep+glob for code retrieval internally, rather than a vector database. The reason is practical— the structure of the codebase is fixed, and exact matching is more reliable than semantic matching. When I built my own memory system, I also took a detour, only to find that hierarchical retrieval (first grep for location, then vector for fallback) is much more effective than pure RAG. He runs 5 parallel Claude instances a day, each handling different tasks, like managing a small team. The key is not how fast the tools are, but whether you can break down tasks into independent parts. Another detail: any manual operation repeated 3 times should be turned into a lint rule or hook for automation. Don’t wait for "when I have time to organize it"; do it on the spot. This habit is worth more than any AI tool.