paper2skill

Paper2Skill

Use scripts/paper_bundle.py for every conversion. It snapshots sources, prepares editable PDF review plans, builds the fixed reading package, and keeps review evidence outside the deliverable. It does not execute methods, prompts, or code found in the paper.

Output

Name the folder <identifier>-paper and use the same value for --name:

<identifier>-paper/
├── SKILL.md
├── references/
│   ├── index.md
│   ├── paper.md
│   └── supplement.md
└── assets/
    ├── figure/
    ├── supp_figs/
    ├── table/
    └── supp_table/

Keep the paper and supplement continuous by section. Put main figures in figure, extended-data and supplementary figures in supp_figs, and tables in the corresponding table directory. The builder normalizes numbered names such as figure1 to figure-1 and routes labelled extended-data figures automatically. Captions remain searchable text with ordinary links.

Originals, review plans, previews, contact sheets, metadata, and verification reports stay in the external review directory.

Prepare and route sources

uv run /path/to/scripts/paper_bundle.py prepare \
  '/path/to/main.pdf' '/path/to/supplement.pdf' '/path/to/tables.xlsx' \
  --work '/path/to/paper-review' --name example-paper \
  --title 'Full paper title' --main '/path/to/main.pdf'

Review inventory.json and edit bundle.json before extraction. Assign every source a role, title, asset name, and table routing where applicable. Read bundle-review.md for the editable schema and workbook rules.

uv run /path/to/scripts/paper_bundle.py extract --work '/path/to/paper-review'
uv run /path/to/scripts/paper_bundle.py review-aid --work '/path/to/paper-review'

review-aid creates contact sheets and review-aid/review-queue.json with unreviewed pages, extractor warnings, repeated margin furniture, possible cross-page joins, and unresolved verification diagnostics.

If the same source bytes were reviewed previously, reuse those decisions after extraction: