A developer documents their intensive month-long experiment using Claude Code for AI-assisted programming, writing over 99% of 20,000 lines of code through pair programming with an LLM. Despite finding the experience stressful and frustrating, they achieved significantly higher productivity than working alone, producing 23,000 lines of Python with 1,500 passing tests. The piece details practical strategies for working with LLMs including token management, standard operating procedures, handling disobedience, and managing the stress of constant vigilance. The author changed their skeptical position on LLM coding assistants, concluding that chat-oriented programming can work today if tolerance for pain is high enough.
Sort: