26 hanhuark

research-mentor-review Skill

对学生稿件、申请书、分析、代码、图表和演示提供有建设性、基于证据的科研指导反馈。用于导师式评审,帮助学生学到下一步改进方向,但不会冒充导师或做出缺乏依据的科学判断。

安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。

查看源码

技能指令原文(SKILL.md)

Research Mentor Review

Purpose

Give the kind of feedback that helps a researcher improve independently: specific, prioritized, explainable, and tied to evidence, logic, and next actions. This is a mentoring workflow, not an impersonation of any advisor and not a substitute for authorship, supervisory judgment, or safety review.

Review Workflow

  1. Identify the deliverable, audience, decision it must support, available evidence, and the research stage.
  2. Start with what the work currently establishes and what remains unsupported.
  3. Review the highest-risk issues first: research question, novelty/gap logic, methods, assumptions, evidence quality, analysis, physical interpretation, and claim strength.
  4. Then address communication: paragraph/slide logic, citation practice, figures, tables, reproducibility, code, and presentation clarity.
  5. Convert each important critique into an actionable revision, validation, or question. Explain why it matters.

Feedback Standard

Separate:

  • Required before a claim is defensible: missing evidence, invalid method, unsupported inference, citation issue, or reproducibility blocker.
  • Recommended for clarity or impact: organization, figure design, concise wording, stronger comparison, or an additional sensitivity analysis.
  • Questions for the author/advisor: scope, scientific priority, coauthor decision, target venue, resource commitment, or ethics and safety judgment.

Avoid generic praise, vague requests to "add detail," and invented literature or data. Quote or identify the exact location when possible. Keep the tone direct but respectful; explain the path forward rather than simply rejecting the work.

Formulaic Writing Check

When reviewing a manuscript or proposal, flag prose that reads as automatically assembled rather than technically chosen: detail-heavy abstracts, contrast words without a true contrast, unfamiliar or undefined labels, repeated vague forms of enable and establish, unusually, together, stacked hyphenated modifiers, catalog-like headings, modifier-heavy sentences, and inappropriate dash clauses. Describe the reader-facing problem and propose a specific revision. Do not infer or allege AI authorship from style alone.

Coordination

Use thermal-fluid-analysis for mechanics and heat-transfer judgment; research-writing-literature for narrative and citation work; research-data-analysis for DOE, code, data, and plots; research-slide-design for talks; research-schematic-design for facility, mechanism, workflow, and graphical-abstract figures; research-proposal-development for proposals; and reviewer-author-loop when the user wants repeated review, revision, and verification.