Duet — AI skill for Claude Code

duet

[](https://github.com/shubharya-os/duet/actions/workflows/ci.yml) [](https://www.python.org/) [](LICENSE) [](https://shubharya-os.github.io/duet/)

**Read in:** English · [简体中文](docs/i18n/zh-CN.md) · [日本語](docs/i18n/ja.md) · [한국어](docs/i18n/ko.md) · [Español](docs/i18n/es.md) · [Português (BR)](docs/i18n/pt-BR.md)

**Two AI models build it. The harness checks the rules.**

**In one sentence:** AI writes most of the code now, and the same AI decides when it is finished. duet makes two AIs from two different companies work the same task and requires both to approve the identical result before it counts as done.

**Why that matters:** an agent grading its own homework will tell you it passed. Two agents that must agree, with your tests run by neither of them, cannot quietly settle for "good enough" — one of them has to be convinced.

**Does it work?** Given one sentence and an empty directory, it built an expense-tracker web app — single-page UI, JSON API, SQLite, CSV export — in **35 minutes**. Then it was attacked by hand, in a real browser as well as over HTTP: SQL injection stored as plain text, stored XSS rendered harmless, CSV formula injection defused, `0.1 + 0.2` totalled to exactly `0.30`. Every run is written down in [docs/QA.md](docs/QA.md) — including the ones where both agents agreed on something broken, and what caught it.

*duet is a developer tool: it runs in a terminal and drives Claude Code and ChatGPT through your existing subscriptions.*

duet build "an expense tracker web app"      # from nothing: criteria, then tests, then code
duet