Research Writing Prompts

Research Writing Prompts

Overview

This skill provides a curated prompt library for academic paper writing workflows. Use it when the user wants either:

1. The raw prompt template itself, ready to copy or adapt.
2. A customized prompt rewritten for the user's current paper/project.
3. Direct execution of the prompt logic, where Codex performs the writing task instead of only returning the prompt.

Quick Start

1. Identify the task type first:
   • translation
   • rewriting / polishing
   • shortening / expansion
   • logic checking
   • reviewer-style critique
   • experiment analysis
   • figure / table naming
   • model selection
2. Read references/prompt-index.md to find the right section.
3. Load only the needed section from references/prompt-library.md.
4. Choose one of three output modes:
   • prompt-only: return the best prompt template.
   • prompt-plus-adaptation: return the source prompt plus a version customized for the user's task or paper.
   • direct-execution: apply the prompt logic internally and return the finished writing/result.

Intent Router

Use this routing table before loading any detailed prompt section.

• If the user says 翻译, 中翻英, 英翻中, 把这段转成英文论文, or equivalent: route to translation prompts.
• If the user says 润色, 重写, 改写, 优化表达, 学术化, or equivalent: route to polish / rewrite prompts.
• If the user says 缩短, 压缩, 扩写, 补充展开: route to shortening / expansion prompts.
• If the user says 逻辑不顺, 帮我检查逻辑, 有没有跳跃, flow: route to logic-check prompts.
• If the user says 像 reviewer 一样看, 审稿, 挑刺, review my paper: route to reviewer-style critique prompts.
• If the user says 去AI味, 更自然, 降低机器感: route to de-AI-tone prompts, but keep claims faithful to the evidence.
• If the user says 实验分析, 帮我分析结果, 写实验部分: route to experiment-analysis prompts.
• If the user says 图标题, 表标题, 画图建议, 架构图: route to figure / table / diagram prompts.
• If the user says 模型选择, 哪个模型更好, 帮我选路线: route to model-selection prompts.

If multiple intents appear together, use this priority:

1. reviewer / logic diagnosis
2. translation
3. rewrite / polish
4. shorten / expand
5. experiment / figure / table support
6. model selection

When the user already provides manuscript text, prefer direct-execution after routing instead of returning a long prompt block.

Task Map

• Translation: Use 中转英, 英转中, or 中转中.
• Rewrite / polish: Use 缩写, 扩写, 表达润色（英文论文）, 表达润色（中文论文）, or the two 去 AI 味 sections as needed.
• Review / diagnosis: Use 逻辑检查 or 论文整体以 Reviewer 视角进行审视.
• Experiment writing: Use 实验分析, 论文架构图, 实验绘图推荐, 生成图的标题, or 生成表的标题.
• Research decision support: Use 模型选择.
• Skill onboarding / examples: Use Skills 的配置, Skills 总览, and 使用场景与示例 Prompt only when the user is asking how to use the repo itself.

Usage Rules

1. Prefer direct execution when the user has already supplied text, figures, tables, or a LaTeX section to edit.
2. Prefer prompt-plus-adaptation when the user says a prompt is useful and wants a reusable task-specific version.
3. Prefer prompt-only when the user explicitly asks for the template itself.
4. Preserve the user's target format:
   • LaTeX for paper source editing.
   • plain text for Word-oriented Chinese writing.
   • bilingual output only when the chosen prompt requires it.
5. If the user already has a fixed paper template or storyline, adapt the prompt to that context instead of returning a generic version unchanged.

For Ongoing Paper Projects

When the user is working on an active paper rather than experimenting with prompts:

1. Extract the useful prompt logic from the library.
2. Translate that logic into direct editing choices on the current manuscript.
3. Keep claims aligned with the user's actual evidence, tables, and figures.
4. Avoid returning large prompt blocks unless the user explicitly wants them.

For benchmark-style papers, prefer prompt logic that emphasizes:

• protocol control,
• claim-evidence alignment,
• clear writing boundaries,
• reviewer-facing justification,
• and distinction between recognition ranking and deployment ranking.

References

• Use references/prompt-index.md as the navigation layer.
• Use references/prompt-library.md as the source prompt library.
• Do not load the entire prompt library unless the user is explicitly asking for a broad survey of all prompts.