shimo4228.hashnode.devI Asked Gemini How Its Own Deep Research Works — Half of It Was InaccurateI brought up token economics with Gemini, and the conversation expanded into a comparison of Deep Research architectures. Gemini eloquently explained the design philosophy differences between itself, ChatGPT, and Claude. It was a fascinating conversa...1d ago·10 min read
shimo4228.hashnode.devBuilding a Claude Code × Kimi K2.5 Hybrid Environment: Opus Designs, K2.5 ImplementsClaude Code (Opus 4.6) is the flagship model in the Claude family, excelling at complex design decisions. However, its per-token pricing can be a concern. Using Opus for routine implementation tasks feels wasteful. That's where Moonshot AI's Kimi K2....2d ago·7 min read
shimo4228.hashnode.devOffloading AI's Weak Spots to Shell Scripts — Designing, Building, and Publishing a Skill Audit CommandIntroduction I built a /skill-stocktake command to periodically audit my Claude Code skill collection. The motivation was simple: let AI judge skill quality. But along the way, I went through four complete redesigns and stumbled into an unexpected di...3d ago·13 min read
shimo4228.hashnode.devThe Real Power of Claude Code Isn't Code Generation — It's Autonomous Context Orchestration"Let AI write your code." That's the first thing people say about Claude Code. It reads files on its own, writes tests, and gets the build to pass. But after a month of daily use, I realized that code generation is not what makes Claude Code valuable...3d ago·7 min read
shimo4228.hashnode.devRunning Claude Code from iPhone via SSH + tmuxWhat Lies Beyond Abandoning GUI In my previous article, I wrote about switching from Cursor to Zed. It was about resolving the discomfort of "a heavy container for light contents" by going with a lighter editor. This time, I go one step further. I di...4d ago·12 min read