← ClaudeAtlas

mathproof-skilllisted

Construct, audit, explain, or research mathematical proofs with correctness-first reasoning. Use when solving proof problems; checking a submitted proof; locating the first unsupported step; finding missing assumptions or counterexamples; explaining a proof pedagogically; researching named, recent, or uncertain theorems; or formalizing and machine-checking a claim, proof, or counterexample in Lean 4 with Mathlib.
Ztang-Yit-Xiaang/mathproof-skill · ★ 1 · AI & Automation · score 74
Install: claude install-skill Ztang-Yit-Xiaang/mathproof-skill
# MathProof Skill Treat plausibility as insufficient. Establish each material step from assumptions or earlier results, and try to falsify the argument before accepting it. The reasoning workflow is informal; Lean 4 with Mathlib is its machine checker. Never claim that an informal audit guarantees correctness, and never call a result machine-checked unless the Lean protocol below passed. ## Select the mode Infer the mode from the request. Ask only if the intended task is genuinely ambiguous. - `PROVE`: construct a proof from a claim and hypotheses. - `AUDIT`: inspect a proposed proof without silently repairing it. - `EXPLAIN`: teach an existing proof while preserving its logic. - `RESEARCH`: locate and verify external theorems, prior proofs, or counterexamples. - `INVESTIGATE`: keep a durable, replayable record of one claim, when the user asks for one. Follow [investigate.md](references/investigate.md). If modes overlap, make `AUDIT` precede `PROVE`: diagnose before repairing. ## Maintain the assumption ledger Before reasoning, record the objects, domains, quantifiers, hypotheses, and target. Track properties that may not be invented silently, including: - nonzero, nonempty, finite, compact, closed, or bounded; - continuous, differentiable, measurable, or integrable; - symmetric, Hermitian, PSD, invertible, full-rank, or diagonalizable; - convex, strictly convex, smooth, or coercive; - independent, identically distributed, centered, or finite-variance. When a step r