adversarial-spec
A Claude Code plugin that refines a spec through a debate between several models until they reach consensus
Medium risk
We rate an entry medium when the tool runs code, makes network calls or reads project files. Check what exactly it does before installing.
Why this level
- Sends the spec to external models from several providers
- External model calls are paid and grow with the number of rounds
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/adversarial-specDetects the agents on your machine, checks the risk and pins the version.
Other ways to install
Run one by one in the Claude Code chat
/plugin marketplace add zscole/adversarial-spec
/plugin install adversarial-spec@adversarial-specInstall the plugin: claude plugin marketplace add zscole/adversarial-spec, then claude plugin install adversarial-spec. Set a key for at least one provider, for example OPENROUTER_API_KEY, and run /adversarial-spec with the product description.
Other ways from the author
claude plugin marketplace add zscole/adversarial-spec
claude plugin install adversarial-specRequires Python 3.10 or newer, the litellm package and one provider key.
This is third-party code. Review the repository files before installing.
What it does
The plugin builds a product spec through adversarial review by several models. Claude drafts a PRD or a technical spec, then other models critique the document in parallel and Claude adds its own critique. Then a loop runs: synthesize the feedback, revise and repeat until all models and Claude agree. After that comes your review period, where you can request changes or extra rounds. The result is a document that has been examined from several angles.
Who it is for. For developers and product managers who want a spec vetted from several angles.
Good fit when
- You want to find gaps and weak spots in a spec
- You want a PRD or tech spec vetted by several models
- You want to put a draft through adversarial review
Not a fit when
- You have no key for any external model provider
- You are not ready for the cost of external model calls
Example request
Review the spec for a Redis rate limiter service through a model debate and drive it to consensusLimitations
It needs Python 3.10 or newer, the litellm package and a key for at least one external model provider. Some providers, such as OpenAI and Anthropic, are not reachable from Russia without a VPN, while others may work, so availability depends on your choice. External model calls are paid, and the cost depends on the number of debate rounds. For enterprise needs it supports routing through AWS Bedrock.
How to disable. Remove the plugin with claude plugin remove adversarial-spec.
Security check
- Sends the spec to external models from several providers
- External model calls are paid and grow with the number of rounds
README in short
The README describes a Claude Code plugin that refines a spec through a debate between several models. Claude drafts, external models and Claude itself critique, feedback is synthesized and the document is rewritten until consensus. Installation is via claude plugin marketplace add and claude plugin install, and it needs Python and the litellm package. It supports many providers via key variables, with an OpenRouter mode and routing through AWS Bedrock. MIT licensed.
FAQ
How many keys do I need?
A key for one provider is enough. Through OpenRouter you can reach several providers with a single key.
When does the debate loop stop?
When all connected models and Claude agree, after which your review period begins.
Related
A skills library that gives coding agents a development process: brainstorming, planning, TDD, subagents and code review
Skills for real engineers by Matt Pocock
Skills For Real Engineers
Small composable skills for engineering with agents: plan grilling, TDD, bug diagnosis, code review and architecture
GitHub toolkit for spec-driven development: the specify CLI adds agent commands and skills to a project, from principles to implementation
Reference MCP servers
Model Context Protocol servers
Official reference MCP servers: Filesystem, Fetch, Git, Memory, Sequential Thinking, Time and Everything