Adversarial Spec

Adversarial Spec is an open-source Claude Code plugin by zscole for writing product specs. Claude drafts the document, sends it to other models for critique, adds its own, revises, and repeats until all of them sign off. The idea is that several models arguing over a spec catch gaps a single reviewer would miss. What you can do with it - Describe a product, or hand it an existing document, and get a PRD or a technical spec back - Start with an optional interview that captures your requirements first - Have models from OpenAI, Google, xAI, Mistral, Groq, Deepseek and others critique each round in parallel, through their APIs, OpenRouter, AWS Bedrock or the Codex and Gemini CLIs - Steer the critique toward security, scalability, performance, UX, reliability or cost, or give the opponent models professional personas - Review the result yourself and run more cycles, see a diff between rounds, resume a saved session, track the cost of each run, and export the final spec as a task list How it connects - In Claude Code: claude plugin marketplace add zscole/adversarial-spec, then claude plugin install adversarial-spec. Run it with /adversarial-spec and a description of what you're building. - It needs Python 3.10 or later, the litellm package, and an API key for at least one other model provider (or a signed-in Codex or Gemini CLI). - The final document is written to spec-output.md or tech-spec-output.md, and can also be sent to Telegram if you set that up. Good to know - Each critique round calls the other providers on your own keys, so runs cost money with those providers; the plugin reports the cost. - MIT licensed.