yhbcode000/paper-share-skills/paper-to-bilibili

Paper to Bilibili

This is the sole full-pipeline orchestration identity. It composes these promoted packages plus the retained managed Rednote workflow:

• paper-to-beamer: paper and MinerU content to compiled 16:10 SUSTech Beamer slides.
• paper-video-cover: optional poster generation for reference; the normal video cover remains the generated cover artifact.
• paper-slides-to-video: physical-page render, narration extraction, TTS, landscape assembly, cover, and Bilibili metadata.
• rednote-video-uploader (managed): required portrait video generation.
• paper-bilibili-uploader: fail-closed validation, dry-run, upload, and durable receipt.

Outputs use $PP_ROOT/论文分享/<VENUE YEAR - SHORT>/ with md_output/, slides-beamer/, and video/.

Modes and phases

slides: paper to compiled slides

When an entry has a nonempty arxiv_id (or the user provided an arXiv link) and no source exists on disk (no root PDF, no paper_src/ with .tex), download the TeX source first:

python \
  "<SKILLS_DIR>/paper-download-arxiv-paper-source/scripts/download_source.py" \
  "<arxiv_id>" --output "<paper_dir>"

The downloader creates <paper_dir>/arXiv-<id>.tar.gz + <paper_dir>/paper_src/; then follow skill://paper-to-beamer Phase 1 onward (Phase 1's move step is skipped since the paper dir already exists). For PDF_ONLY output, copy the PDF to the paper dir root (the pdf_downloaded gate requires a root PDF).

Follow skill://paper-to-beamer for content authoring. Required gates before continuing:

1. slides-beamer/main.tex contains the paper-specific narrative from paper information and motivation through methods, experiments, limitations, takeaways, references, Q&A, and 3–8 useful appendix slides.
2. The document uses the SUSTech 16:10 template (aspectratio=1610).
3. slides-beamer/main.pdf is nonempty, with no undefined controls or severe overflow.

The batch CLI only scans and persists disk evidence for this agent-authored phase; it does not fabricate slides.

Narration authoring gate

Copy the compiled TeX to video/main_with_narration.tex and add exactly one nonempty % NARRATION: entry per physical PDF page. Narration is polished Simplified Chinese with English technical terms inline. It expands the technical argument rather than reading bullets, avoids visual-only references such as “本页” or “如图所示”, and covers appendix pages too.

The transition from slides_done to narrations_done occurs only when both the compiled slides and nonempty annotated TeX exist.

video: narrated landscape and portrait artifacts

python <SKILLS_DIR>/paper-to-bilibili/scripts/batch.py video "<PAPERS_JSON>"

For each narrations_done paper, the orchestrator calls the sibling paper-slides-to-video full command. After child success it rescans and atomically persists video_done. It then calls the managed rednote-video-uploader/scripts/portrait_video.py, rescans, validates cover and Bilibili metadata, and atomically persists upload_ready. Child failures propagate and never advance state.

Quality gates are exact frame/text/MP3 cardinality, nonempty narration and audio for every physical page, a nonempty landscape MP4, a nonempty portrait MP4, a nonempty cover, and schema-valid Bilibili metadata.

bilibili: validate or upload

Always review a dry-run first:

python <SKILLS_DIR>/paper-to-bilibili/scripts/batch.py bilibili "<PAPERS_JSON>" --dry-run

Dry-run delegates to the promoted uploader and must leave upload_ready unchanged. Remove --dry-run only after review. There is no default upload timeout; an operator may opt in with --upload-timeout SECONDS.

Normal uploads require P1 landscape then P2 portrait. Metadata uses singular tag, a positive tid, valid copyright/source fields, a structured description, and a nonempty cover. A paper becomes uploaded only when video/upload_result.json is valid and contains a matching BV URL plus ISO-8601 timestamp.

State contract

batch.py rescans the disk with this exact precedence:

1. valid upload_result.json → uploaded
2. nonempty landscape + nonempty portrait + cover + schema-valid video_meta.json → upload_ready
3. nonempty landscape → video_done
4. compiled slides + nonempty annotated TeX → narrations_done
5. compiled slides → slides_done
6. MinerU Markdown → mineru_done
7. source PDF or extracted TeX source (paper_src/ containing .tex) → pdf_downloaded
8. otherwise → pending

Every successful child command is followed by a rescan and atomic replacement of the input JSON. State changes are printed as STATE <paper>: <old> -> <new>. Dry-run validates and delegates but never invents a new state. A failed child preserves the last grounded state and its exit code is propagated.

Batch input

papers.json is an array of objects with a nonempty absolute dir_path; dir_name, arxiv_id, pdf_path, and descriptive fields may accompany it. The orchestrator records _status in each object.

Prerequisites

• Python 3.10+ (if python is not on PATH, use py -3).
• Dependencies are declared in pyproject.toml (currently none at runtime); install the project with uv sync --project <SKILLS_DIR>/paper-to-bilibili when adding any.
• <SKILLS_DIR> is the directory containing the installed skill folders (e.g. ~/.claude/skills, ~/.codex/skills, .omp/skills); sibling skills (paper-slides-to-video, paper-bilibili-uploader, paper-download-arxiv-paper-source) must be installed next to this one so sibling resolution stays project-local and deterministic.
• PAPER_PYTHON (optional): interpreter used to launch sibling CLIs; defaults to the interpreter running batch.py.

Related skills

• skill://paper-bilibili-uploader
• skill://paper-download-arxiv-paper-source
• skill://paper-slides-to-video
• skill://paper-to-beamer
• skill://paper-video-cover
• skill://rednote-video-uploader