| name | conference-papers-read2md |
| description | Use when Codex needs to read a conference paper from data/papers/*.json and generate a standalone Obsidian-compatible deep-reading Markdown note under data/_inbox, with online images by default and optional localization only when needed. Trigger for requests like "Use $conference-papers-read2md 精读 data/papers/xxx.json", "根据 paper JSON 生成 Obsidian md", or "read conference paper JSON to Markdown"; this skill is independent from the conference-papers website note JSON and dist build. |
Conference Papers Read2MD
Generate dailypaper-style Obsidian Markdown notes from data/papers/*.json. This is a standalone Markdown export workflow: do not edit paper["note"], do not modify data/papers/*.json, do not run website validation/build scripts, and do not treat the result as site data.
This skill adapts huangkiki/dailypaper-skills paper-reader resources. Preserve its deep-reading style and quality bar; only the input source and output location differ.
Workflow
- Resolve the site root and target paper JSON.
- Accept an explicit JSON path, paper id, or title substring.
- Run the context helper from this skill directory:
python3 scripts/paper_json_context.py data/papers/PAPER_ID.json
-
Read the helper JSON output and create the inbox directory:
output_dir: normally data/_inbox/{paper-id}
assets_dir: normally data/_inbox/{paper-id}/assets, used only if images are explicitly localized
- final note path:
data/_inbox/{paper-id}/{MethodName}.md
- use
suggested_method_name as a starting point, but correct it if the paper clearly uses a different method/model name.
-
Locate the paper source in this priority order:
- Existing
arxiv_id, arxiv_url, or pdf_url from the JSON.
- If missing, search arXiv using
arxiv_search_query from paper_json_context.py.
- Then try
url, detail_url, or project_url.
- Stop with a clear missing-source message if no full paper, arXiv HTML/PDF, or usable project/source page can be found.
-
Read the full paper before writing:
- Prefer arXiv HTML because it exposes figures and equations.
- Use PDF when HTML is unavailable.
- Check the abstract, method sections, experiments, appendix if relevant, all figures, all tables, and important formulas.
- Use
references/cv-dl-terminology.md only when terminology translation or naming is uncertain.
-
Write the Markdown note using assets/paper-note-template.md.
- Fill YAML frontmatter with paper JSON metadata where available.
- Fill the metadata table with authors, venue as
{Venue} {Year}, and categories/topics from the paper JSON.
- Use the existing dailypaper note style: Chinese technical deep reading, not a short summary.
- Keep inline
[[概念]] links for important technical terms, datasets, methods, simulators, losses, and architectures.
- If no Zotero collection exists, set
zotero_collection: _inbox or a topic-derived path from the paper JSON.
-
Use online images by default.
- Default image strategy: online images by default.
- 默认保留 arXiv HTML / 项目主页图片 URL, using normal Markdown image links such as
.
- Do not run
scripts/localize_note_images.py by default.
scripts/localize_note_images.py is an optional localization fallback only when image URLs are unreachable, the user explicitly asks for offline archiving, or the note must be migrated to an offline Obsidian vault.
- When optional localization is needed, run:
python3 scripts/localize_note_images.py "data/_inbox/{paper-id}/{MethodName}.md"
- The script stores images under the note's
assets/ directory and rewrites links as Obsidian wikilinks such as ![[assets/MethodName_fig1.png|600]].
- If optional localization fails, keep the best reachable network link or explain which figure needs manual recovery.
- Self-check the finished note.
- Read
references/quality-standards.md before finalizing.
- Read
references/image-troubleshooting.md when arXiv HTML image paths, project images, or PDF extraction are involved.
- Use
references/concept-categories.md to choose good concept link names; do not create concept notes or MOCs unless the user explicitly asks.
Quality Requirements
- Include every figure from the paper. Do not include only teaser images.
- Figures may use network image URLs by default; local image files are not a completion requirement.
- Include every important formula with LaTeX, meaning, and symbol explanations.
- Include every important table with complete rows/columns or a faithful Markdown transcription.
- Explain why each figure/table/formula matters; do not paste captions as detached dumps.
- Preserve the dailypaper sections: metadata, one-sentence summary, core contributions, background, method details with formulas, figures/tables, experiments, critical thinking, and quick card.
- Use
[[概念]] links naturally in the body. Keep them useful for Obsidian, even though this skill only writes the paper note and images.
- Never write a skeleton note in place of full reading. If the paper cannot be read completely, leave a clear failure report instead of a low-quality
.md.
Important Non-Goals
- Do not modify
data/papers/*.json.
- Do not generate or validate
paper["note"].
- Do not call
../conference-papers/scripts/build_site.py, validate_notes.py, or resolve_links.py.
- Do not refresh Obsidian MOCs or run git automation.
- Do not batch-read all candidates by default; process the user-specified paper unless they explicitly ask for a bounded batch.
Bundled Resources
assets/paper-note-template.md: dailypaper Obsidian note template.
scripts/paper_json_context.py: normalize a paper JSON into source/output context.
scripts/localize_note_images.py: optional localization fallback to download note images into inbox assets/ and rewrite links when explicitly invoked.
references/quality-standards.md: detailed figure/formula/table quality bar.
references/image-troubleshooting.md: arXiv/project/PDF image fallback guidance.
references/concept-categories.md: concept naming and category guidance for [[概念]] links.