Summarize Paper
Overview
Use this skill to summarize one academic paper from the user-provided source. Extract the paper's purpose, main contributions, techniques, limitations, and future prospects, then output both a Markdown summary and an Excel workbook when the environment can create files.
Core Rules
- Base every factual claim on the supplied paper. Do not add external background, author intent, impact claims, or field context unless the user explicitly asks.
- Separate paper-stated content from inference. Mark each item as
原文明确,原文概括,合理推测, or未提及. - Use
合理推测only when the paper does not directly state the item but the inference follows from nearby methods, results, limitations, or conclusion text. - Keep inferred language cautious, using phrases such as
可推测,可能,倾向于说明, or未来可进一步. - Do not over-interpret. If evidence is weak, mark the item as
未提及or explain the uncertainty in the evidence column. - Preserve important details the paper actually states, including task setting, dataset, model/framework, experimental setup, evaluation metrics, and named techniques when relevant.
- Use short evidence references instead of long quotations. Prefer page, section, table, figure, or paragraph anchors when available.
- If the paper text is incomplete, scanned, or extraction quality is poor, state that limitation before the summary and avoid filling gaps with confident claims.
- In user-facing Markdown, group content by the six required dimensions and show multiple claims as bullet points under the same dimension. Do not present repeated
研究目的,主要贡献, or other repeated dimensions as separate visual sections. - Keep the structured JSON/Excel rows at claim level for traceability, but ensure the rows are ordered by dimension so downstream pages can group them cleanly.
Workflow
- Identify the paper source type: PDF, DOCX, Markdown, plaintext, pasted excerpt, or extracted text.
- Read the abstract, introduction, method, experiments/results, discussion, limitations, and conclusion first.
- Scan the remaining body for claims that affect the required categories.
- Build a claim list with evidence anchors before drafting the final summary.
- Fill the required dimensions:
研究目的主要贡献使用技术/方法实验与结果不足/局限未来前景/后续工作
- For each dimension, include all material points explicitly present in the paper.
- Add cautious inferred items only after all original-content items, and label them clearly as
合理推测. - Produce the Markdown output first, grouping the claim list under dimension headings.
- Create the Excel output using
scripts/write_paper_summary_excel.pywhen file creation is available. - Archive the generated Markdown, JSON, and Excel files using
scripts/archive_summary_outputs.pywhen file creation is available, so the literature management webpage can auto-discover the new paper. Prefer a local structure where one paper folder contains the source PDF/text extraction and the generated summary outputs; the webpage will only import summary output files.
Markdown Output
Use this structure:
# 论文总结:<论文标题或用户提供文件名>
## 基本信息
- 标题:
- 作者:
- 年份/会议或期刊:
- 研究领域:
- 资料完整性说明:
## 总览
<用 3-6 句话概括论文解决的问题、方法和结论。只写原文可支持的信息。>
## 逐项总结
### 研究目的
- 【原文明确|高】...(依据:Abstract; Introduction)
- 【原文概括|高】...(依据:Introduction)
### 主要贡献
- 【原文明确|高】...(依据:Contribution bullets)
### 使用技术/方法
- 【原文明确|高】...(依据:Method)
### 实验与结果
- 【原文明确|高】...(依据:Experiments; Table 1)
### 不足/局限
- 【原文明确|高】...(依据:Limitations; Conclusion)
### 未来前景/后续工作
- 【原文明确|高】...(依据:Conclusion)
## 推测内容清单
| 推测项 | 推测依据 | 置信度 | 后期核查建议 |
|---|---|---|---|
| ... | ... | 高/中/低 | ... |
## 需注意的原文限制
- ...
Guidelines:
- Keep original and inferred content visually distinct by putting the type and confidence at the beginning of each bullet, for example
【原文明确|高】. - Each required dimension should appear once as a heading. Put multiple points under that heading as bullets.
- If a dimension has no support in the paper, include one bullet labeled
【未提及|】and explain that the paper does not provide the information. - Add a separate
合理推测bullet only if a careful inference is useful, and also list it in推测内容清单. - Include the
推测内容清单even when empty; write无when no inference was used. - The Markdown grouped bullets and the JSON/Excel claim rows must describe the same underlying claims, even though Markdown is grouped for readability.
Excel Output
When the user asks for Excel output, or when files can be created, produce an .xlsx workbook in addition to Markdown.
Use one worksheet named 论文总结 with columns:
维度类型总结原文依据/推测依据置信度后期核查建议
Use scripts/write_paper_summary_excel.py to create the workbook from a JSON file:
python scripts/write_paper_summary_excel.py summary.json paper_summary.xlsx
Expected JSON shape:
{
"paper_title": "Paper title",
"authors": "Author A; Author B",
"venue": "Journal or Conference; Year",
"year": "2026",
"field": "Research field",
"overview": "A short paper-level overview for the literature manager card.",
"rows": [
{
"dimension": "研究目的",
"basis_type": "原文明确",
"summary": "What the paper states.",
"evidence": "Abstract; Introduction.",
"confidence": "高",
"review_suggestion": ""
}
]
}
Keep JSON and Excel rows as claim-level records, ordered by the six required dimensions. Include the paper-level JSON fields above when available so the literature management webpage can render compact cards with title, venue/year, topic, and short overview instead of inferring them from claim rows.
Library Archive Output
After creating Markdown, JSON, and Excel outputs, archive them for the dynamic literature management webpage:
python scripts/archive_summary_outputs.py summary.json paper_summary.md paper_summary.xlsx
Archive behavior:
- If
SUMMARIZE_PAPER_LIBRARY_DIRis set, archive into that directory. - Otherwise archive into
~/Documents/summarize-paper-library/inboxwhen the Documents folder exists. - Otherwise archive into
~/.summarize-paper-library/inbox. - Put each paper in a stable paper-title subfolder and keep
summary.json,summary.md,summary.xlsx, andmanifest.jsontogether. - Use
--folder-name "Author - Year - Title"when you want the folder name to match a manually curatedpaper/<paper folder>/...library. - Use
--timestampedonly when versioned runs should be kept in separate folders.
The webpage can monitor the archive inbox or the parent paper folder and automatically refresh when new skill outputs appear. It recursively scans subfolders and imports only summary.*, paper_summary.*, and *_paper_summary.* files, ignoring PDFs and extracted text files.
Quality Check
Before finishing:
- Confirm each required dimension is present exactly once as a Markdown heading.
- Confirm multiple points inside a dimension are bullets under that heading, not repeated dimension blocks.
- Confirm every factual summary item has an evidence anchor.
- Confirm every inferred item is labeled
合理推测. - Confirm no inferred item is mixed into an
原文明确or原文概括row. - Confirm the JSON and Excel claim rows reconcile with the grouped Markdown bullets.
- Confirm missing source information is marked as
未提及instead of invented. - Confirm the generated files were archived for the literature management webpage when file creation is available.