菜单
Pinned-revision API reconnaissance before proposing any Lean theorem. Classifies each requirement as direct, local lemma, or infrastructure blocker against the exact pinned mathlib commit, and emits a verdict plus a manifest instead of wrapper theorems.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
基于 SOC 职业分类
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/utensil/formal-land --skill ut-lean-recon命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Design of a Lean formalization slice before writing any code: the five-question compact design check, convention locks by definitional acceptance tests, the authoritative-spec principle, and the characteristic-API rules for public declarations. Slice selection (the gates and the operational algorithm) is ut-lean-roadmap's concern.
Run and verify a Lean project at the toolchain level: slice worktrees, safe fork rebases, pinned Lean and mathlib, cache-first builds with cache reuse, no-sorry and axiom audits, fresh-checkout reproduction, and the verification discipline for executable Lean content.
Shorten Lean proofs at the mathematical interface by replacing locally rebuilt machinery with the library abstraction that already names the object. Survey the pinned revision, search by structure before writing lemmas, state at natural generality, extract shared criteria on first reuse, and land structural API at the project's permitted boundary before specializing.
| name | ut-lean-recon |
| description | Pinned-revision API reconnaissance before proposing any Lean theorem. Classifies each requirement as direct, local lemma, or infrastructure blocker against the exact pinned mathlib commit, and emits a verdict plus a manifest instead of wrapper theorems. |
Reconnaissance is the exploration step that runs before any Lean theorem is proposed. Given a target statement and the project's pinned mathlib revision, it answers one question: which parts of the target already exist, which can be built from what exists, and which need new infrastructure? The answer is a verdict plus a manifest. Design, golf, and review consume that output, so recon must be complete and reproducible before any plan depends on an interface. The main direction is recon before design; design may re-invoke recon in focused form (a probe against the pinned revision) when its convention locks or signature choices raise new API questions, and the two iterate until the design's contracts are pinned against recon evidence.
The pinned mathlib commit, the Lean toolchain, and the project commit are part of every claim. Grep the pinned checkout, not the live master, unless the survey explicitly covers mathlib pull-request history. Record the checkout path in the manifest. The pinned checkout is the project's Lake-managed mathlib under .lake/packages/; it is not re-cloned. Surveys that need mathlib source history read the local reference checkout, pulled once per slice (skills README, Reference repositories), never a fresh clone.
Grep the pinned checkout for the mathematical structure and the types involved (a group type, a quotient construction, a bilinear form, a contraction operation), not only for a hoped-for theorem name. Names and docstrings drift; structure does not. If the search finds a type that is a definitional re-spelling of an existing quotient or construction, unify with the existing type instead of adding a third spelling.
For every candidate declaration, read the statement, assumptions, universe levels, and namespace. A name is not evidence. Record what the declaration actually states, including hidden typeclass assumptions.
For each requirement of the target, record one of:
Build the source-role versus existing-declaration table in MANIFEST.md, one row per requirement, each row tied to the pinned revision.
If a source-level definition is already covered by mathlib, the conclusion is a documented no-gap result, not a wrapper theorem. A wrapper that merely restates an existing declaration duplicates library work and is rejected in review. Record the covered boundary precisely: hypotheses, direction, conventions. If the difference from the library version is substantive (a different definition shape, a different convention), record that difference instead of papering over it.
Before proposing anything new, search mathlib, the project's merged history, and its open pull requests for the same or adjacent declarations. A collision means contribute to or build on the existing work, not re-open it.
When an authoritative milestone names sibling endpoints, census that family and record shared machinery, its earliest natural owner, and each endpoint's first consumer.
No equality is claimed until a convention table records, for each side: the direction of the map, the source and target forms, and the signs or scalar factors. A change-of-form map, a quotient identification, or a normalization constant must be matched explicitly. Record the table in the manifest.
Remembered names are unreliable. Every probe that fails is journaled with the wrong remembered name and the correct route. Plans must never depend on an invented interface; the journal is the evidence that they do not.
Every claim in the manifest carries one label:
Probes live in a scratch directory outside every repository and are never committed. A cache-backed build imports the pinned library and elaborates the candidate signatures or the minimal concrete instance. A spike may use a named sorry boundary only to test interface shape; a successful elaboration is not evidence that the postponed theorem is easy.
The recon output is:
Attach MANIFEST.md with the tables and journal, and hand the verdict to design, golf, or review.