meleantonio/prove-that-shit

Prove That Shit is a plugin that helps AI coding agents attempt open mathematical problems, the kind of unsolved conjectures associated with mathematician Paul Erdős. It packages a specific workflow described by Shouqiao Wang, built around careful problem selection, precise prompting, long autonomous research runs, and adversarial review of any resulting proof. The goal is not to instantly solve famous problems but to give agents a structured methodology for trying. The plugin works across multiple coding agents, including Cursor, Claude Code, Gemini CLI, GitHub Copilot, and any tool that reads an AGENTS.md file. You install it by cloning the repository into the right directory for your chosen agent. Each agent picks up the same set of shared skills, commands, and rules through adapter files tailored to that tool. The workflow has several steps. You start by selecting a sharply stated open problem, either manually or through a command that filters for tractable candidates. You then craft a problem-specific prompt that defines exactly what counts as a valid solution and explicitly bans weak or partial results. That prompt becomes the goal of a long autonomous run. The README notes that hard instances in the original workflow took six to thirty-two hours. After a run completes, an independent agent audits the proof before you treat it as solved. An optional step formalizes the result in Lean, a proof verification language. The repository organizes its methodology into shared skills, agent definitions, slash commands, and templates. Commands like /prove, /select-problem, /craft-prompt, and /audit-proof map to stages in the pipeline. The templates encode the attempt, audit, and repair loop that the workflow relies on. The README is candid about limitations. This is a methodology pack, not a system that automatically solves open problems. Human mathematicians should still audit important claims. Community acceptance on sites like erdosproblems.com remains separate from an agent declaring a problem resolved locally.