← ClaudeAtlas

duetlisted

Hand a task to Claude Code and ChatGPT together, so they build it and check each other until both sign off. Use when the user asks for a second opinion from another model, a cross-check, "what would ChatGPT say", or wants two agents to work a task jointly. Also use when the user is about to ship something risky and wants an independent reviewer in the loop.
shubharya-os/claude-chatgpt-duet · ★ 3 · AI & Automation · score 76
Install: claude install-skill shubharya-os/claude-chatgpt-duet
# duet — a second model in the loop You are Claude Code. `duet` lets you bring ChatGPT in: either for a one-shot review of what you just did, or for a full build where the two of you take turns until you both sign off. The point is that a single agent grades its own homework. duet puts something in the loop that is allowed to say no, and gives that objection weight. ## Running it ```bash {{DUET}} run "the task" --gate "pytest -q" ``` Both agents take turns in the workspace until they agree. It never ends on one agent's say-so: both must vote DONE on the same workspace state, with no blocking objection open and the gate passing. To have the other model check work that already exists, give it that as the task: ```bash {{DUET}} run "review the current diff against the task; fix what is actually wrong" \ --gate "pytest -q" --rounds 4 ``` ## Rules for you - **Always pass `--gate`** with the project's real test or build command. It is what keeps "done" honest — the harness runs it and it overrides both agents' opinions. A session without a gate is two models agreeing by argument alone. - **Never run `duet run` on uncommitted work the user cares about** without saying so first. Both agents edit the workspace directly. Suggest a branch. - **Report the findings, not the exit code.** The user wants to know what ChatGPT objected to and whether it was right — say which findings you agree with and which you would push back on. You are allowed to disagree with the revi