wtf is modelduet?

ModelDuet is a tool for developers who use AI coding assistants. It pairs two different AI models so that each one handles the part of a coding task it is best at, rather than having a single model do everything. The idea is that when one model both plans and checks its own work, it can miss its own mistakes. By splitting the work, ModelDuet gives you a second pair of eyes on every change. The tool works as a manually invoked skill inside Claude Code, a command line AI assistant from Anthropic. It coordinates two specific models: Fable 5 and GPT-5.6 Sol. Fable handles the discovery, planning, review, and verification steps. Sol handles the actual implementation and bug fixing. The two trade off in a loop that ends when the plan passes review or after five rounds, whichever comes first. The workflow starts with a user request. Fable first explores the codebase and writes a plan. Sol then implements that plan. Fable reviews the result and runs verification checks. If Fable finds blocking issues, Sol fixes them and Fable reviews again. If everything passes, Fable produces a final report. Each run starts from a clean Git state, creates an isolated branch for the task, and never merges or pushes code on its own. To install ModelDuet, you clone its repository into your Claude Code skills directory. You need both Claude Code and the Codex CLI already installed and authenticated, plus Git. Usage is as simple as typing a slash command followed by your task description. The tool is intentionally manual, so Claude will not trigger it without you asking. Safety is a priority. If your working directory has uncommitted changes, the run stops rather than risking those changes. Fable reads the full diff and runs verification commands. Security flaws, failed checks, and unmet criteria all block approval. The project is MIT licensed and is an independent effort not affiliated with Anthropic or OpenAI.