Harper Reed: LLM codegen workflow

Harper Reed describes a workflow for building software with LLM code generation, split into greenfield projects and incremental work on existing codebases. For greenfield work, he uses a conversational model such as ChatGPT 4o or o3 to hone an idea one question at a time into a developer-ready spec saved as spec.md, then a reasoning model such as o1, o3 or r1 turns the spec into a step-by-step prompt plan and a todo.md checklist. Execution is done by pasting those prompts into claude.ai or into aider, with tests run after each step. For existing code, he packs the repository into an output.txt with repomix and uses his LLM command-line tool with mise tasks to generate missing tests, reviews and READMEs. He says the planning step takes about 15 minutes.