Summary Papers

Summary-Papers

Use this skill when the user wants to batch-read PDF papers from one local folder and export paper summaries as Markdown, Excel, or both.

Required user input

Ask for or confirm these values before doing any work:

paper_root: /absolute/path/to/paper/folder
output_root: /absolute/path/to/output/folder
output: markdown | excel | all

Interpret the input strictly:

• paper_root is one local folder.
• Recursively scan paper_root and all subfolders for .pdf files.
• Ignore non-PDF files.
• output_root is where final Markdown, Excel, and Markdown assets are written.
• Intermediate JSON and reports are written under /.summary-papers/.
• output controls final artifacts:
• markdown: write one Markdown file.
• excel: write one single-sheet .xlsx file.
• all: write both.

Important environment rule

Only claim that files were written when the runtime truly has filesystem access to the requested local paths.

• If the environment can access the local paths, write files directly.
• If the environment cannot access the user's local filesystem, do not pretend the write succeeded. Explain the limitation and ask for accessible files or generate content as a deliverable.

Workflow

Follow this sequence.

1. Validate absolute paper_root and output_root.
2. Run dependency check:

python3 scripts/check_deps.py check

This checks both the current Python environment and the default dependency target /tmp/summary_papers_deps.

1. If dependencies are missing and the user approves installation, install them into a temporary target:

python3 scripts/check_deps.py install --target /tmp/summary_papers_deps

Then invoke later commands with:

PYTHONPATH=/tmp/summary_papers_deps python3 scripts/paper_summary_cli.py ...

1. Prepare the run:

• always scan recursively;
• extract text for every PDF;
• extract Markdown image assets only when output is markdown or all;
• write manifest.json, extracted_text.json, run_report.json, and summary_draft.json under /.summary-papers/.

python3 scripts/paper_summary_cli.py prepare "/path/to/papers" --output-root "/path/to/output" --assets

Omit --assets when output is excel.

1. Read /.summary-papers/extracted_text.json and summarize every paper into the complete structured JSON schema below.
2. Save the completed structured summaries as JSON, usually based on /.summary-papers/summary_draft.json.
3. Export the final artifact:

python3 scripts/paper_summary_cli.py export "/path/to/output/completed_summaries.json" --output-root "/path/to/output" --output all

Complete Structured JSON Schema

Create one object per paper. Preserve source fields and include these keys:

• title
• authors
• year
• venue
• publish+time
• keywords
• github_url
• 研究现状
• motivation
• insight
• 核心贡献
• method
• 实验结论
• 局限性
• 其它
• zotero_link
• pdf_link
• 源文件
• 源路径
• relative_path
• title_author_image
• framework_image
• status
• error

Use empty strings for genuinely unavailable fields unless the Markdown field rule below says to render 未找到 or 未提供.

Language and Common Field Rules

Summary fields must be written in Chinese. Keep title and publish+time in the original English when the PDF provides English metadata. For keywords, use the paper's original English keywords when explicitly provided; if the paper has no explicit keywords, summarize exactly three English keywords or short phrases from the paper content.

Markdown and Excel share the same common summary fields and the same target lengths:

• 研究现状: about 100 Chinese characters.
• motivation: about 100 Chinese characters.
• insight: about 100 Chinese characters.
• 核心贡献: about 200 Chinese characters.
• method: about 500 Chinese characters.
• 实验结论: about 200 Chinese characters.
• 局限性: about 200 Chinese characters.
• 其它: about 200 Chinese characters.

Excel Output Rules

Excel is always one workbook with one sheet named papers.

Columns:

• 源文件
• 源路径
• title
• publish+time
• keywords
• 研究现状
• motivation
• insight
• 核心贡献
• method
• 实验结论
• 局限性
• 其它

Field rules:

• 源文件 and 源路径 are required traceability columns.
• title: English paper title from the PDF when available.
• publish+time: only information stated in the PDF. Prefer Venue Year, for example AAAI 2024.
• keywords: use explicit English keywords from the paper when available. If the paper has no explicit keywords, summarize exactly three English keywords or short phrases yourself. Separate keywords with semicolons.
• Common summary fields must follow the target lengths in Language and Common Field Rules.

Do not embed images in Excel. Do not add image-path columns unless the user explicitly asks.

Markdown Output Rules

Markdown is always one .md file containing all processed papers.

For each paper, include:

• 标题
• 作者
• 年份
• 发表的期刊/会议
• keywords
• GitHub 链接
• PDF 链接
• Zotero 链接
• title/authors image if extracted
• method/framework image if extracted
• 研究现状
• motivation
• insight
• 核心贡献
• 方法
• 实验结论
• 局限性
• 其它

Use 未找到 for unavailable bibliographic or content fields. Use 未提供 for Zotero/PDF links only when no useful link/path is available.

Image Rules

Only extract and use images for Markdown output.

• Always include the title/authors image if extraction succeeded.
• Include the method/framework image only if extraction succeeded and looks plausibly relevant to the paper's method.
• Do not fabricate or hallucinate an image.
• If a framework image is missing, omit that image block.
• Prefer one framework image per paper.
• When uncertain, prefer omitting the framework image over using a likely wrong teaser, result chart, or first-page illustration.

Image filenames must be ASCII-safe and manifest-driven:

assets/20260701_143015_paper_001_safe_stem__title.png
assets/20260701_143015_paper_001_safe_stem__framework.png

The leading YYYYMMDD_HHMMSS timestamp is generated at scan time. Do not use arbitrary paper titles or original Unicode filenames as generated image filenames.

Run Reports

Every run must preserve machine-readable status files under /.summary-papers/:

• manifest.json: recursive PDF inventory.
• extracted_text.json: inventory plus extracted text and title guesses.
• run_report.json: per-paper status, errors, text extraction status, and asset extraction status.
• summary_draft.json: blank complete-schema draft that can be filled by the agent.

Keep final user-facing outputs in output_root:

• paper_summaries.md
• paper_summaries.xlsx
• assets/

For each paper, record failures conservatively instead of stopping the whole batch whenever possible.

Quality Bar

• Base every field only on the PDF itself unless the user explicitly asks for external enrichment.
• Do not use web search by default.
• Prefer empty strings over guesses in structured JSON.
• Do not copy long sentences from the paper. Compress into concise, high-density summaries.
• Keep Markdown and Excel common summary fields aligned in content and target length.
• Keep the output language Chinese except title, publish+time, and keywords. Preserve English originals when available; for missing keywords, write exactly three English keywords or short phrases yourself.
• Mention uncertainty when metadata, text extraction, or figure extraction is ambiguous.

Example

请使用 Summary-Papers 处理这个目录，递归总结所有 PDF，并同时输出 Markdown 和 Excel。

paper_root: /Users/me/Documents/papers
output_root: /Users/me/Documents/paper-output
output: all

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: WindZh03
• Source: WindZh03/Summary-Paper-Skill
• License: MIT

Install and usage instructions live in the source repository linked above.