菜单
read a survey paper by building a taxonomy, unified notation table, representative-method comparison, and math-first deep dives. use when the user wants 阅读综述, taxonomy extraction, 统一符号表, 代表方法对比, benchmark summaries, or a reusable survey study template.
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。
基于 SOC 职业分类
align a research paper with its codebase by scanning the repo, locating configs and training scripts, mapping paper modules to code modules, and auditing implementation or reproduction gaps. use when the user wants 结合代码读论文, 论文代码对齐, implementation verification, repo scanning, or reproducibility analysis.
deconstruct a research paper into reusable components, generate component cards and dependency graphs, explain one component in depth, and produce a full structured obsidian-friendly paper breakdown. use when the user wants 论文组件化解构, 单组件精讲, 全文结构化拆解, componentization, block-library extraction, or reverse-engineering of author choices.
audit a paper like a strong reviewer with claim-evidence-falsify tables, stepwise red-teaming, and comparison against field consensus or active controversy. use when the user wants 强批判审稿报告, 分步红队, 共识与争议对比, rigorous criticism, reviewer-style questions, or evidence-disciplined paper auditing.
turn a paper or repo into a practical research extension plan with baseline reproduction, boundary tests, mechanism hypotheses, ablation priorities, improvement routes, milestones, and writing outline. use when the user wants 找边界, 思考可做的 idea, boundary testing, ablation planning, or to move from reading to concrete experiments.
generate, filter, and rank research ideas from a broad direction using a codex-native idea funnel modeled on the aris idea-discovery workflow. use when the user wants literature mapping, 8-12 candidate ideas, feasibility and novelty filtering, devil's-advocate review, optional pilot planning or pilot execution, and an idea_report-style output that preserves both recommended and eliminated ideas.
| name | survey-taxonomy-reader |
| description | read a survey paper by building a taxonomy, unified notation table, representative-method comparison, and math-first deep dives. use when the user wants 阅读综述, taxonomy extraction, 统一符号表, 代表方法对比, benchmark summaries, or a reusable survey study template. |
默认输出语言：中文。默认把综述读成“taxonomy + 统一符号 + 代表方法对比”，而不是逐段复述。
当用户在读 survey / review / tutorial 类论文，想快速建立领域结构，再对代表方法做数学或机制层面的精读时，使用这个 skill。
它合并两种稳定模式：
guided: 按章节带读，每次只讲一个部分，适合交互式学习full-map: 一次性输出 taxonomy、符号表、代表方法表、benchmarks、挑战与未来方向math-first: 背景最少化，集中在统一形式化、损失函数、推导和统计假设未指定时默认 full-map。
整理出：
缺信息时：
【未知】先说明：
输出 taxonomy 树或 Mermaid 思维导图，并解释：
优先使用 references/taxonomy-template.md。
统一符号表，至少整理：
优先使用 references/notation-table-template.md。
从每个大类挑 2-3 个代表方法，比较：
使用 references/method-comparison-template.md。
如果用户偏数学，重点做：
独立公式用 LaTeX 单独成行，不放进代码块。
总结：
给出：
如果用户要求“一次只讲一个部分”，按以下顺序：
每次只输出一个部分，结尾问是否进入下一部分。