Math Olympiad Adversarial Solver
Solve and stress-test IMO, Putnam, USAMO, and AIME problems with adversarial proof verification.
The skill explicitly restricts the standard workflow to reasoning, forbids web access, and requires honest abstention when gaps remain, reducing data and external-service risk. However, tool restrictions rely on prompting rather than enforced isolation; deep mode permits Bash/Python, and agent launches and PDF compilation lack user confirmation, data-flow disclosure, or rollback guidance, so points are deducted.
The staged workflow, retry policy, early-exit voting, and calibrated abstention are reasonably clear. However, the stated 8–12 solver agents conflict with model-tier defaults, “36 results” is not stable against those configurations, tool restrictions are not enforceable, and no executable tests or concrete diagnostic handling for abnormal inputs are provided; the score is therefore capped and reduced under static review.
Audience, competition-math scenarios, input types, output formats, and trigger phrases are fairly explicit, with positive and negative trigger examples. Boundaries remain broad around terms such as “competition math”; non-fit cases for general mathematics, teaching, or research are not defined, and Chinese-language output or mainland-China environment support is not documented. The core function does not depend on overseas network services, but it does depend on agent orchestration capabilities.
The skill has metadata, progressive workflow sections, model-tier guidance, reference prompts, trigger evaluations, and limitation notes; repository context also supplies Apache-2.0 licensing, official-maintenance provenance, and CI maintenance signals. Still, the skill has no skill-level changelog, named maintainer, update path, installation/dependency notes, or complete troubleshooting guidance, and its references inconsistently describe the number of verifier patterns, so points are deducted.
The design directly targets solving, proof checking, counterexamples, batch problems, and presentation, while calibrated abstention can improve practical usefulness. Static materials contain no real outputs, test results, or independently reproducible execution evidence, and agent counts, context isolation, and PDF compilation depend on an unverified runtime; light-review usability therefore cannot be established and the score is reduced within the static cap.
The files provide a fairly detailed workflow, reusable prompt templates, trigger evaluations, and a cited research background, giving some auditability. They provide no committed tests covering key skill paths, real CI execution evidence for this skill, or independently reproducible third-party results, and the research citation lacks verifiable details in the supplied material; the score therefore does not exceed the static-review ceiling.
- Tool restrictions are primarily prompt-based; subagents may still have access to the full tool set, and deep mode permits local code execution.
- Solver counts, agent labels, model-tier defaults, and verifier-pattern counts are inconsistent, which may affect result association and cost control.
- There are no skill-level execution tests, failure fixtures, named maintainer, or changelog; PDF output also depends on LaTeX tools being available.
- Chinese-language behavior and specialized handling of Chinese competition problems are not documented.
What it does & when to use it
Math Olympiad is a skill in Anthropic’s official Claude Code Plugins Directory for competition mathematics. It interprets the statement, generates multiple candidate solutions, and verifies cleaned proofs in fresh contexts using specific failure patterns. The workflow can revise proofs, switch to bounded local computation in deep mode, and format verified results for presentation. When the evidence is insufficient, it explicitly returns “no confident solution” instead of bluffing.
Checks alternative interpretations of a problem and selects the intended reading; solves AIME-style numeric problems through multiple attempts and majority voting; generates candidate proofs for IMO, Putnam, and USAMO problems; removes solver traces before verification; launches fresh-context adversarial verifiers to test theorem hypotheses, hidden gaps, circular reasoning, invalid generalizations, and divergent series; starts revision cycles when holes are found; uses bounded Bash/Python computation in deep mode when required; and can compile a PDF with pdflatex after verification if the environment supports it.
- An IMO, USAMO, or Putnam participant needs a rigorous proof checked for hidden gaps.
- An AIME solver wants several independent attempts and a majority vote on the numeric answer.
- A student or coach wants an existing competition proof attacked rather than merely paraphrased.
- A researcher suspects a proposed lemma is false and wants small counterexamples or a proof audit.
- A user is processing a complete competition set and wants one independently verified result per problem.
Pros & cons
- Uses interpretation checks and named adversarial failure patterns tailored to competition mathematics.
- Separates solver reasoning from verifier context to reduce agreement bias.
- Uses asymmetric voting thresholds and pigeonhole-style early exits.
- Optimizes for calibrated correctness rather than forcing an answer.
- Supports per-problem processing and PDF compilation for complete competition sets.
- The standard tight-budget workflow relies on reasoning rather than computation.
- Deep mode permits only bounded local Bash/Python computation and still forbids web access.
- PDF generation depends on pdflatex and the referenced compilation scripts being available.
- The source provides no test suite, platform matrix, or empirical accuracy guarantee.
- The full workflow can require many solver, verifier, and revision agents.
How to install
The skill is located at plugins/math-olympiad/skills/math-olympiad/SKILL.md. The README documents Claude Code marketplace installation with the format /plugin install {plugin-name}@claude-plugins-official; the supplied source does not identify the exact published plugin name for this skill, so the placeholder should not be guessed. The repository is licensed under Apache-2.0.
How to use
Submit a problem or proof-review request such as Solve this IMO problem: [problem], Is this proof correct? [proof], Find a counterexample to [claim], or Prove this olympiad inequality: [inequality]. Proof problems use candidate generation, context-isolated verification, and revision when necessary; numeric problems use majority voting. The skill may return no confident solution when verification does not close the gaps.