夜雨聆风
一个现成主 skill + 一组自己维护的小 skill的路线
Codex 现在支持基于 `SKILL.md` 的 Agent Skills，Claude Code 也支持同类 `SKILL.md` skill；为了两边通用，尽量只用标准字段 `name` 和 `description`，少用某一家的扩展字段。
Codex 官方说明 skill 是包含 `SKILL.md`、可选 `scripts/`、`references/`、`assets/` 的目录，并支持显式或隐式调用；Claude Code 也支持把 skill 放到 `.claude/skills/<skill-name>/SKILL.md` 或 `~/.claude/skills/<skill-name>/SKILL.md`，可自动触发或用 `/skill-name` 调用。
最推荐的技能
1. Master-cai/Research-Paper-Writing-Skills
用途包括 Abstract、Introduction、Method、Experiments、Conclusion 的起草/重写，段落流畅性、章节逻辑、claim-evidence 对齐，以及从审稿人视角做 pre-submission self-review。仓库也明确给了 Codex 和 Claude Code 的安装方式。
2. imbad0202/academic-research-skills
这个更重，像一套“科研流水线”：deep research、academic paper、academic reviewer、academic pipeline 等，还包含 LaTeX hardening、citation check、revision coaching、style calibration 等功能。适合你想让 Claude Code 参与从调研、写作、审稿模拟到修改的完整流程
3. K-Dense-AI/scientific-agent-skills
这个不是单纯论文写作 skill，而是科学研究技能库，覆盖大量科学计算、数据库和分析工作流。它更适合做实验分析、可视化、科学工具调用的补充，不是写作主力。仓库称其兼容支持 open Agent Skills 标准的工具，包括 Claude Code 和 Codex。
建议维护的轻量 skill 组合
不要只做一个巨大的 `research-paper-writing`。科研写作更适合拆成 5 个小 skill：
你可以先建这个：
```md
---name: research-paper-writing
description: Improve ML/CV/biomedical imaging research papers. Use for rewriting Abstract, Introduction, Related Work, Method, Experiments, Conclusion, checking claim-evidence alignment, improving academic logic, reviewer-perspective self-review, and LaTeX paper structure.
Research Paper Writing Skill
You help write and revise research papers in ML, computer vision, medical imaging, and 3D/4D reconstruction.
Core principles
1. Preserve the author's technical meaning. Do not invent results, baselines, datasets, metrics, citations, or conclusions.
2. Improve logic before language. First check the argument structure, then improve wording.
3. Every strong claim must be backed by one of: experiment result, ablation, figure/table, theorem, citation, or clearly stated assumption.
4. Prefer precise academic prose over vague hype.
5. For biomedical or clinical claims, avoid overstating clinical validity unless supported by evidence.
Workflow
When revising a section:
1. Identify the section type: Abstract, Introduction, Related Work, Method, Experiments, Discussion, Conclusion, or Rebuttal.
2. Summarize the intended message in 1-3 bullets.
3. Detect missing logic:
- unclear problem motivation
- weak contribution statement
- unsupported claim
- missing comparison to prior work
- experiment/result mismatch
- vague limitation
4. Rewrite the text.
5. Provide a short change log explaining what was improved.
Output format
Return:
1. Revised text
2. Key improvements
3. Remaining risks or missing evidence
放置位置建议：
txt
Codex 项目级
.agents/skills/research-paper-writing/SKILL.md
Claude Code 项目级
.claude/skills/research-paper-writing/SKILL.md
Codex 官方推荐项目内使用 `.agents/skills`，用户级使用 `$HOME/.agents/skills`；Claude Code 项目级使用 `.claude/skills/<skill-name>/SKILL.md`，个人全局使用 `~/.claude/skills/<skill-name>/SKILL.md`
可以先装 Master-cai/Research-Paper-Writing-Skills，然后自己补三个小skill：`claim-evidence-check`、`latex-paper-maintainer`、`reviewer-response`。这样比装一个超大的“科研全流程智能体”更稳，也更容易在 Codex 和 Claude Code 之间保持一致。
