科研 Skill

独立论文同行评审

paper-review

独立多视角论文评审与方法学审计

官方收录发布者 local安全审核通过v0.1.0更新于 2026-08-06

Skill 说明

面向学术论文的独立同行评审 Skill;支持单一完整评审或方法、统计、创新性与可复现性多视角 panel,按用户选择的重点和目标检查研究设计、统计有效性、逻辑、图表、写作与复现风险,并给出分级、可执行的审稿意见。 | 输入:PDF、DOCX、Markdown 或 LaTeX 论文,可选目标期刊、稿件类型和修订状态 | 输出:Synopsis、Critical/Major/Minor 问题、参考依据、修改优先级和编辑建议

SKILL.md

Academic Manuscript Review

Routes a manuscript to an independent, fresh-context reviewer that evaluates it for methodological soundness, statistical validity, logical consistency, and reproducibility, and returns a structured peer review. This skill is a thin dispatcher: it decides how to run the reviewer and in which mode. The review procedure, checklists, statistical and figure guides, principles, and output template all live in references/ and are loaded by the reviewer, not duplicated here.

When to use

Activate when the user wants peer-review feedback on a manuscript (journal article, conference paper, preprint).

Accepted manuscript inputs

Accept a manuscript supplied as Markdown (.md), Word (.docx), or PDF (.pdf); LaTeX (.tex) remains supported. Determine the intake branch from the supplied file rather than requiring a particular format. The detailed conversion and inspection rules are in references/review-procedure.md.

Why a fresh-context reviewer

Review validity depends on independence: a reviewer that shares the conversation that produced the manuscript is biased toward it. Run the reviewer in a separate context and pass only framing (manuscript path, target journal, manuscript type, revision status), never the authoring rationale. This is why the reviewer is a subagent on tools that support one, and an inline procedure where they do not.

Review preflight (ask the user before every review)

Before reading, converting, or dispatching the manuscript, collect the three review settings below. Do not silently apply the defaults for an interactive review request.

  1. Use the host's structured ask-user capability. Prefer request_user_input when it is available; otherwise use the runtime's structured ask_user tool. Ask these three questions in one interaction when the tool supports multiple questions:
    • Mode: Single complete review or Panel multi-perspective review.
    • Focus: Comprehensive, Methods and statistics, Figures and reproducibility, or Writing and logic.
    • Goal: Strict pre-submission assessment or Revision-oriented suggestions.
  2. If neither structured tool is available, ask the same three questions in chat and stop the current turn until the user answers. Do not begin the review or choose a default on the user's behalf.
  3. Record the three answers in the reviewer framing together with manuscript path, target journal, manuscript type, and revision status. If the user has explicitly supplied one or more of these settings in their request, retain those values and ask only for the missing settings.

Apply the selected focus as the weighting of the review, not as permission to ignore validity-threatening issues outside that focus. Comprehensive uses the full procedure. The other focus choices prioritize their named checks and keep the remaining passes proportionate. Strict pre-submission assessment uses the existing high-bar severity calibration. Revision-oriented suggestions keeps the same evidence and severity standards, but prioritizes actionable fixes and clear revision sequencing.

Modes

  • Single complete review: one independent reviewer applies the full procedure end to end.
  • Panel multi-perspective review: spawn independent reviewers in parallel on complementary lenses, then a synthesis pass. Lenses: methods/design, statistics, and novelty/significance (add reproducibility for methods-heavy or hardware papers). Each reviewer reads the whole manuscript but weights its lens and scores independently from references/; a final synthesis pass merges them into one Critical/Major/Minor review and surfaces genuine disagreement rather than averaging it away.

Dispatch

Run Review preflight first. Then pick the branch for the current tool. In every branch the reviewer follows references/review-procedure.md, using the selected mode, focus, and goal as reviewer framing.

  • Claude Code: Task(subagent_type: "paper-review", ...) passing the manuscript path, target journal/type, mode, focus, and goal. For panel mode, launch one Task per lens in parallel, then a final synthesis Task.
  • Codex CLI: plugin installation exposes this skill, not a Codex subagent. To use a fresh-context Codex reviewer, first copy agents/templates/paper-review.toml to ~/.codex/agents/ or .codex/agents/, then invoke that configured agent if the current Codex surface supports /agent. For panel mode, ensure max_threads covers the lens count. If no Codex subagent is configured or available, use the fallback branch.
  • Copilot CLI: plugin installation exposes this skill and, through .github/plugin/plugin.json, the .agent.md reviewer in agents/templates/. Invoke that configured agent when the current Copilot surface supports custom agents; use /fleet for panel mode when available. If running outside a plugin install, copy agents/templates/paper-review.agent.md to .github/agents/ or ~/.copilot/agents/. If no custom agent is available, use the fallback branch.
  • Fallback (no subagent support, or the user wants an interactive in-thread review): first locate the rubric ($CLAUDE_PLUGIN_ROOT/skills/paper-review/references, else find . -type d -path '*/skills/paper-review/references' | head -1); if it cannot be found, stop and tell the user to install the manuscript plugin rather than reviewing from memory. Then follow references/review-procedure.md directly in this context.

The brain (do not duplicate into dispatch or agent shells)

  • references/review-procedure.md -- step-by-step procedure: intake, read, methodology, logic, literature, reproducibility, figures, writing, output.
  • references/methodology-checklist.md, references/statistical-review-guide.md, references/figure-review-guide.md -- the assessment checklists and guides.
  • references/review-principles.md -- review philosophy and severity calibration.
  • references/review-output-template.md -- the Synopsis / Critical / Major / Minor / References / Editor Note format.
  • examples/sample-manuscript-review.md -- worked review for calibration; examples/sample-manuscript-excerpt.md -- sample manuscript input for testing.
  • Sister skill manuscript:humanizer -- AI-writing patterns to flag in the prose-quality pass.

常见问题

独立论文同行评审能做什么?

面向学术论文的独立同行评审 Skill;支持单一完整评审或方法、统计、创新性与可复现性多视角 panel,按用户选择的重点和目标检查研究设计、统计有效性、逻辑、图表、写作与复现风险,并给出分级、可执行的审稿意见。 | 输入:PDF、DOCX、Markdown 或 LaTeX 论文,可选目标期刊、稿件类型和修订状态 | 输出:Synopsis、Critical/Major/Minor 问题、参考依据、修改优先级和编辑建议。

如何安装独立论文同行评审?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「独立论文同行评审」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

独立论文同行评审需要哪些权限?

该科研 Skill声明需要:本地文件。安装前会再次列出,由你确认后才会生效。

独立论文同行评审经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。