What Makes a Good Tool for Claude Code (lalitm.com)

A Claude Code user reflects on why some tools work remarkably well with LLM-based tool calling even when they’re complex, arguing that the Unix “do one thing well” philosophy is not the whole answer. From their experience with git—sprawling, stateful, and counterintuitively handled effortlessly by Claude—they distill three hallmarks that make a tool LLM-friendly: long-lived or widely used tools (massive presence in training data), excellent documentation (built-in help, man pages, external docs), and informative error messages that give corrective suggestions. Examples include classic Unix utilities, git, npm, Docker, Beancount (whose strong docs let Claude learn usage patterns), and the Rust compiler (whose helpful diagnostics let the model iteratively correct mistakes). For the AI/ML community, the takeaway is practical: LLMs typically leverage memorized examples and published guidance rather than fresh reasoning, so tool authors should prioritize discoverability and feedback over minimalism to maximize LLM utility. Clear docs, sample workflows, and precise, suggestion-rich errors create a feedback loop that lets models use even complex, stateful systems reliably. This reframes how we design integrations for LLMs—favor understandability and diagnostic signals to unlock large gains in tool-calling performance.