Paper Deep Read（论文十问精读）
Overview
This skill turns a research paper into a structured, citable reading report. It applies a fixed ten-question framework so that every deep read follows the same disciplined process: objective restatement of the paper (Q1–Q5, Q8), plus explicit critical judgment from the reader (Q6, Q7, Q9). The output is grounded in the paper's own structure — every claim points to a section, figure, or table — and never invents content.
When To Use
Activate this skill when all of the following hold:
- The user provides a paper (local PDF, a paper URL/DOI, or pasted full text), and
- The user asks to 精读 / 深度阅读 / 总结 / 评价 / 写阅读笔记 on that paper.
Do not activate for: casual one-line summaries, citation formatting, proofreading, or translation of a paper (those are separate tasks). If the user only wants a one-paragraph TL;DR, still run the full framework but present a short summary first, then the full ten answers on request.
Workflow
Step 1 — Ingest the paper
- If the input is a local PDF, read the full document (use the PDF reading capability; do not skim). Extract title, authors, year, venue, and the section/figure/table structure.
- If the input is a URL or DOI, fetch and read the full text first; if only the abstract is available, state that the deep read is limited and proceed with what is accessible.
- If the input is pasted text, use it as-is and note any missing parts.
- Record the paper's structural map (Abstract → Introduction → Related Work → Method → Experiments → Results → Discussion → Conclusion → References; list of Figures/Tables) so every later claim can cite a location.
Step 2 — Read for evidence, not impression
- Read the entire paper before answering. Map each of the ten questions to the section(s) that answer it.
- Distinguish "what the paper says" (objective, for Q1–Q5 and Q8) from "your own judgment" (subjective, for Q6, Q7, Q9).
- Flag anything you cannot verify as "不确定", and anything the paper does not cover as "论文中未提及".
Step 3 — Answer the ten questions
Follow the fixed framework below, in order, by number. The full text of the questions and the answer rules are authoritative — reproduce them verbatim in the report.
> # 论文十问精读
> 请仔细阅读论文全文，然后逐一回答以下10个问题：
> 1. 论文试图解决什么问题？
> 2. 这是否是一个新的问题？为什么？
> 3. 论文的核心思想是什么？
> 4. 论文的方法和实验设计如何？
> 5. 论文的实验结果和数据分析是什么？
> 6. 论文的创新点和贡献是什么？
> 7. 论文的局限性和不足之处是什么？
> 8. 论文的相关工作和前人研究如何？
> 9. 论文的结论是否合理和可靠？为什么？
> 10. 论文的未来研究方向和拓展价值是什么？
>
> ## 回答要求
> - 按编号逐条回答，第1-5、8题客观转述论文内容即可；第6、7、9题要给出你自己的判断。
> - 每题的依据注明来自论文哪个部分（章节/图/表）；论文没提到的写"论文中未提及"，不要编造。
> - 拿不准的写"不确定"。
> - 用简体中文回答。
Step 4 — Emit the structured report
- Use the Markdown template in
references/template.mdas the output skeleton. Fill every question block; do not drop any of the ten. - Lead with a one-line TL;DR, then the ten numbered answers, then a short "判断小结" (your overall verdict on contribution + reliability + whether to read further).
- Keep citations precise: e.g. "（见 Section 3.2）", "（Fig. 4）", "（Table 2）". Never fabricate a section/figure number.
Honesty Rules (non-negotiable)
- Cite the source location for every non-trivial claim; if none exists, write "论文中未提及".
- Use "不确定" whenever the evidence is weak or contradictory; never overstate.
- Do not paraphrase the authors' claims as your own in Q1–Q5 / Q8; reserve judgment for Q6, Q7, Q9.
- Respond in 简体中文 throughout.
Resources
references/
template.md— The fill-in Markdown skeleton for the structured ten-question report. Load it when producing the final output.
