After hundreds of hours using ChatGPT, Claude, Claude Code, and Codex, I noticed I kept running into the same problem. The agents could usually write the code. The harder part was making sure they understood: what the task actually was what systems were risky what was out of scope what success looked like I found myself recreating the same project context, ticket documents, constraints, and acceptance criteria over and over. Eventually I turned the workflow into an open-sourc...