math-olympiad

What it is

math-olympiad is a Claude Code plugin from Anthropic for solving competition math problems from contests such as IMO, Putnam and USAMO. It is built around one problem: self-verification gets fooled. The README says a verifier that can see the reasoning tends to agree with it. It cites arXiv:2503.21934 ("Proof or Bluff"), which found that 85.7% self-verified IMO success drops to under 5% under human grading. The plugin addresses this with context-isolated verification. The verifier sees only the clean proof and never the reasoning trace. The checks are also adversarial and armed with specific failure patterns. Instead of asking "is this correct?", a verifier asks questions like "does this accidentally prove RH?" or "extract the general lemma, find a 2×2 counterexample." The plugin description says these fresh-context verifiers attack proofs with specific failure patterns. The plugin uses calibrated abstention. When it cannot reach a verified answer, it says "no confident solution" rather than bluffing. Once a proof passes verification, a presentation pass produces clean LaTeX/PDF output. For validation, the README reports 17 of 18 IMO and Putnam 2025 problems solved, with 0 false positives and 2 novel proofs found. It points to the skill's eval data in the Anthropic monorepo.

Who it's for

• People working on competition math problems such as IMO, Putnam or USAMO
• Users who want proofs checked by adversarial verifiers that don't see the original reasoning
• Anyone who needs to verify an existing proof
• Users who prefer an honest "no confident solution" over a bluffed answer
• Users who want verified solutions as clean LaTeX/PDF

Requirements

Setup

1. Install the plugin

   Install math-olympiad from the official Claude plugins marketplace with the plugin install command.

   bashCopy

   /plugin install math-olympiad@claude-plugins-official

Examples

> Solve this IMO problem: [statement]