A developer shares a first-hand account of building Cure, a new programming language with dependent types and SMT verification targeting the BEAM, using AI agents as co-developers. After an initial failed attempt, they rebuilt from scratch with a disciplined approach: feeding the AI their own handcrafted code as style examples, building a custom AST-based RAG tool, enforcing strict linting and testing gates after each step, and keeping tasks small and unambiguous. Key lessons include: avoid motivational prompt preambles, ensure tasks have exactly one solution path, review the AI's output stream in real time, and never delegate tasks you couldn't implement yourself — the AI accelerates execution but cannot substitute for architectural understanding.
Sort: