wtf is kimchi?

Kimchi is a Claude Code and Codex plugin that turns a rough idea into documents an AI coding assistant can actually build from. The problem it targets is familiar: you type a vague prompt into an AI assistant, it fills in gaps you never specified, and hours later you end up with something close to what you wanted but not quite right, because the real plan only ever existed in your head. Instead of one agreeable assistant, kimchi runs your idea past eleven expert personas, covering product, architecture, design, security, finance, operations, and more, including one deliberately skeptical staff engineer character. Each persona questions your idea in its own area and pushes back rather than praising it, until most of the hard questions about the idea have real answers. Once the grilling is done, kimchi compiles the results into a set of build ready documents: one file per feature area with user stories, priority order, and business importance, plus a single file that locks in decisions like the tech stack, architecture, exact API contracts, and security and deployment choices, so those do not get re argued later. It also produces an execute file that lets Claude Code or Codex build the project piece by piece across multiple separate sessions, picking up where it left off using tracking files. Before building starts, an audit step checks the documents for problems while they are still easy to fix. You can install kimchi as a Claude Code plugin that updates itself automatically, or clone the repository and run an install script that detects which coding assistant you have and copies the skill files into the right folder for Claude Code or Codex. After installing, you invoke it with a short command and can either run the whole guided process or call up a single persona by name. The project is released under the MIT license.