Paper2wechat

Paper2WeChat

Execute this workflow when producing a Chinese WeChat article from an Arxiv paper.

Inputs

Accept one of:

• Arxiv URL: https://arxiv.org/abs/2301.00000
• Arxiv ID: 2301.00000
• Local PDF path: ./paper.pdf

Optionally accept:

• user preferred style (optional override)
• max length
• max images
• output path

Step 1: Parse Paper And Extract Figures

Run:

bash .agents/skills/paper2wechat/scripts/fetch_paper.sh "<url_or_id_or_pdf>" ".paper2wechat"

Expect output lines like:

• Parsed cache: .paper2wechat/<paper_id>/parsed/<paper_id>.json
• Images dir: .paper2wechat/<paper_id>/images

While running, the parser prints progress logs to stderr (for example download progress and extraction stages). For very large PDFs (default: ≥30MB or ≥50 pages), TeX/source fetching may be auto-skipped to avoid long downloads; override with --source always.

Image extraction behavior:

• For arXiv URL/ID: prefer TeX source images first, then fallback to PDF caption-based extraction.
• For local PDF: use PDF extraction path directly.

Verify parser output before writing:

• title, authors, affiliations, abstract
• sections
• images with url and caption

If no images are extracted, continue with text-only article and state that figures were unavailable.

Step 2: Generate Style Evidence For Agent Decision

Use parsed JSON to generate style evidence from paper content (title, abstract, sections, captions):

python .agents/skills/paper2wechat/scripts/detect_style.py ".paper2wechat/<paper_id>/parsed/<paper_id>.json" --json

Rules:

• Treat script output as evidence, not final style lock.
• If user explicitly requires a style, user preference overrides all.
• If confidence_band is high, usually adopt top candidate.
• If confidence_band is medium or low, let Agent choose from top-2 or use hybrid style.
• If top-2 are close, hybrid style is allowed (for example academic-tech + academic-applied).

See references/style-guide.md for interpretation rules.

Step 3: Build A Practical Summary

Produce a practical summary section before long-form explanation. Ensure it answers all items below:

• 论文解决了什么问题
• 方法的核心创新是什么
• 关键结果指标是什么（优先写具体数字）
• 读者可以直接借鉴的做法是什么
• 落地边界和风险是什么

Keep this summary scannable with 4-6 bullets.

Step 4: Generate The Article

In this skill, article rewriting is done by the Agent directly from parsed JSON.

Use references/article-template.md as the output scaffold and adapt tone by chosen style. Template is a baseline, not a rigid format: adjust section names/order by paper type and audience. Extract useful links directly during writing from parsed JSON text:

• open-source repo links (GitHub/GitLab/HuggingFace, if present)
• related papers/resources for further reading

About 扩展阅读（相关研究 + 技术工具/资源）:

• Content should be grounded in the paper as much as possible: prefer items explicitly mentioned in the paper body (related work, baselines, benchmarks/datasets, toolkits, project pages).
• When possible, add clickable links (arXiv / conference page / project site / GitHub / HuggingFace).
• If the paper does not provide a link and you cannot reliably identify one, it is OK to omit the link (do not guess).

For image links:

• default output file is under .paper2wechat/<paper_id>/outputs/
• use relative image paths like ../images/<image_file>
• image filename must come from .paper2wechat/<paper_id>/parsed/<paper_id>.json (images[].url) instead of guessed naming patterns
• never guess filenames like page_*.png when parsed JSON provides src_*.png.

For image count:

• Do not hard-code to 1-2 images.
• Default to dynamic range 2-6 when images are available.
• Select by relevance and narrative fit: overview/framework first, then method details, then key results.
• Avoid near-duplicate figures or too many small/local patches from one big figure.
• If user sets max images, obey it.

Always include:

• 论文信息块（标题/作者/机构/论文链接/发布日期/开源地址；优先使用 affiliations，缺失时写“未明确注明”，开源地址无则写“未提供”）
• concise 导读
• practical summary section
• method/result sections with context
• 扩展阅读（相关研究 + 技术工具/资源）
• 关键词 hashtag line

Style-aware structure guidance:

• academic-science: emphasize assumptions, experiment setup, reproducibility limits.
• academic-tech: emphasize architecture, implementation details, engineering trade-offs.
• academic-trend: emphasize direction shifts, ecosystem implications, future outlook.
• academic-applied: emphasize use cases, rollout constraints, KPI/ROI.

Step 4.5: Persist Markdown File

After drafting content, write the final markdown to file.

Default output path:

• .paper2wechat/<paper_id>/outputs/<paper_id>.md

Rules:

• Unless user explicitly asks for chat-only output, always write/update the markdown file.
• Ensure output directory exists before writing.
• Do not only return content in the chat pane.
• After writing, report the absolute or repo-relative file path in the response.

Never add tool-credit disclaimers like “本文由...自动生成”. Never output tool-wrapper artifacts in markdown body, including:

• </content>
• <parameter name="filePath">...
• local absolute paths like /Users/...

Step 5: Validate Output Quality

Check the final markdown file:

• Structure is complete and readable on mobile.
• Style and tone match requested audience.
• Dynamic 2-6 images are inserted with contextual text and captions when available and appropriate.
• Claims and numbers align with source content.
• Output path exists (default: .paper2wechat/<paper_id>/outputs/<paper_id>.md).
• MUST verify every image link resolves from output file directory.
• MUST verify markdown does not contain tool-wrapper artifacts (</content>, <parameter name="filePath">, absolute local paths).

Resources

• scripts/fetch_paper.sh: standalone entrypoint to parse paper and extract figures into cache.
• scripts/parse_paper.py: standalone parser with caption-aware figure extraction and fallback strategies.
• scripts/detect_style.py: recommend style from parsed paper content.
• references/style-guide.md: style selection rules and wording constraints.
• references/article-template.md: reusable WeChat article scaffold.

Converted and distributed by TomeVault — claim your Tome and manage your conversions.