| name | paper-share-post |
| description | Create reusable Chinese AI paper-sharing deliverables from a research paper PDF for native Chinese readers. Use when Codex is asked to process a 论文 PDF for 论文分享、自媒体帖子、博客、paper reading notes, or WeChat/Xiaohongshu-style sharing, including extracting/cropping the paper homepage, figures, and tables, then writing a concise Chinese Markdown post with source, link, paper title, research background, contributions, method, and one-sentence summary. |
Paper Share Post
Convert one research-paper PDF into a date-archived share-ready bundle: paper screenshots plus a Chinese Markdown sharing post.
The target audience's native language is Chinese. Write for Chinese-speaking AI readers using natural Chinese phrasing, not translated English sentence structure.
Reference Loading
- Read
references/post-style.md before drafting or revising the Markdown post.
- Read
references/screenshot-spec.md before producing or checking screenshots.
- Use
scripts/pdf_screenshot_helper.py when PDF rendering or image cropping can be automated.
Inputs
Require the paper PDF. Extract metadata from the PDF first, then ask only for missing information that cannot be inferred safely:
- paper source or venue, such as
AAAI 2026, IJCAI 2026, arXiv, or a project page
- canonical paper URL
- output folder, if the user has a specific location
- screenshot scope, if the user wants fewer than all main-text figures and tables
Do not invent source, link, title, task setting, method name, experiment result, or claimed contribution. If a detail is uncertain, mark it as unresolved and ask before finalizing.
Output Layout
Create a per-paper output folder unless the user specifies one. The default root is the current working directory. Use the current date unless the user specifies a date. Derive paper-name from the paper title after metadata extraction, using a readable filesystem-safe short name:
YYYY/
└── MM/
└── MM-DD-论文名/
├── screenshot/
│ ├── 0 文章第一页PDF截图.png
│ ├── 1.png
│ ├── 2.png
│ └── ...
└── 论文分享.md
Example for a paper processed on 2026-06-01:
2026/
└── 06/
└── 06-01-FedALT/
├── screenshot/
│ ├── 0 文章第一页PDF截图.png
│ ├── 1.png
│ └── 2.png
└── 论文分享.md
Do not use the old <paper-name>-share/ or screenshots/ layout unless the user explicitly asks for it. Use screenshot/ exactly, singular.
The final paths must resolve to:
YYYY/MM/MM-DD-论文名/screenshot/
YYYY/MM/MM-DD-论文名/论文分享.md
For example:
2026/06/06-01-论文名/screenshot/
2026/06/06-01-论文名/论文分享.md
Inside screenshot/:
screenshot/
│ ├── 0 文章第一页PDF截图.png
│ ├── 1.png
│ ├── 2.png
│ └── ...
Use 0 文章第一页PDF截图.png for the full first page. Number figure and table screenshots as 1.png, 2.png, ... in the order they should appear in the share post, normally the order of appearance in the main paper.
Workflow
-
Inspect the PDF.
- Identify title, authors if useful, source or venue, URL, abstract, introduction problem, method section, contributions, key experiments, conclusion, and all main-text figures/tables.
- Prefer the paper's own wording for method and contribution names, then rewrite in accessible Chinese.
- Decide the output folder name only after identifying the title. Keep it short, stable, and readable; replace path separators and unsafe characters with
-.
-
Create the output folder.
- Default path format is
YYYY/MM/MM-DD-论文名/.
- Put all images in
YYYY/MM/MM-DD-论文名/screenshot/.
- Put the Markdown post at
YYYY/MM/MM-DD-论文名/论文分享.md.
- If the user gives an output root, create the
YYYY/MM/MM-DD-论文名/ tree under that root.
-
Produce screenshots.
- Capture the PDF first page as
0 文章第一页PDF截图.png.
- Crop every main-text figure and table unless the user narrows scope.
- Keep captions when they help interpret the figure/table; otherwise crop tightly around the visual content.
- Use readable resolution, no cut-off labels, no unrelated column text, and consistent margins.
- If automatic figure detection is unreliable, render relevant pages and crop manually by pixel coordinates.
-
Draft 论文分享.md.
- Follow
references/post-style.md exactly for structure, tone, length, and title constraints.
- Keep the whole post within 900 Chinese characters.
- Use natural Chinese, light emoji, and an AI blogger voice: clear, practical, and easy to scan for Chinese readers.
- Explain the research background first, then contributions, then method, then a one-sentence takeaway.
- Avoid evaluation hype that the paper does not support.
-
Verify before final response.
- Confirm the title starts with
论文分享: and is no more than 20 characters when feasible.
- Confirm all metadata placeholders are filled or explicitly listed as unresolved.
- Confirm the Markdown has the required sections in order.
- Confirm the Markdown is saved exactly as
YYYY/MM/MM-DD-论文名/论文分享.md.
- Confirm screenshots are saved exactly under
YYYY/MM/MM-DD-论文名/screenshot/.
- Confirm screenshot filenames match the required convention.
- Confirm the post is under 900 Chinese characters.
Screenshot Helper
Use the bundled helper for repeatable rendering/cropping:
python scripts/pdf_screenshot_helper.py init paper.pdf out --date 2026-06-01 --title "论文名"
python scripts/pdf_screenshot_helper.py first paper.pdf out/2026/06/06-01-论文名/screenshot
python scripts/pdf_screenshot_helper.py render paper.pdf out/rendered --pages 1,3-5 --dpi 220
python scripts/pdf_screenshot_helper.py crop crops.json out/2026/06/06-01-论文名/screenshot
init creates YYYY/MM/MM-DD-论文名/screenshot/ and prints the paper folder, screenshot folder, and Markdown path. If --title is omitted, it tries to infer a title from PDF metadata or the first page text; verify the inferred folder name before finalizing.
crop expects image pixel coordinates:
[
{
"source": "out/rendered/page-003.png",
"box": [120, 240, 1420, 980],
"output": "1.png",
"padding": 12
}
]
Coordinates are [left, top, right, bottom] in pixels. Resolve relative source paths from the crop JSON file location.
Final Response
Return the absolute paths to:
- the Markdown post
- the screenshot folder
- any unresolved metadata or screenshots that need user confirmation
Keep the final response concise. Do not paste the full post unless the user asks to preview it in chat.