Paper2read

Paper2Read

Create one calm reading surface from a dense paper. Preserve the paper's text, make its structure and argument easier to follow, and keep machine plumbing out of the human-facing report.

All paths below are relative to this Skill root. Keep the environment and paper outputs in the user's active workspace, not in the installed Skill directory.

Local PDF to reading document

If dependencies are unavailable, prepare the workspace environment:

python scripts/setup_environment.py --use-uv --install

Run the complete pipeline:

.paper2read/.venv/bin/python scripts/process_paper.py PAPER.pdf \
  --output paper-output \
  --ui-language zh

The user-facing result is paper-output/report.md. Internal source, interpretation, metadata, and publishing data belongs under paper-output/.paper2read/ and must not be presented as reading material.

Read references/reading-format.md when changing the report or judging its quality.

Non-negotiable reading boundary

• Preserve the complete extracted source; do not replace it with a summary.
• Keep H1/H2/H3, lists, tables, formulas, figures, captions, and reading order where extraction supports them. Crop captioned figures into images/ and place them beside their captions; do not silently reduce a figure to its caption.
• Render the three reading modules as H1, separate them with dividers, and infer numbered paper hierarchy (3 → H2, 3.1 → H3, 3.1.1 → H4).
• Highlight research roles and explicit connectors without rewriting the author's claims. For connector categories, change foreground text color only; keep the background unchanged and never add colored-square swatches.
• If a research question or hypothesis is not explicit, say 原文未明确陈述 or Not explicitly stated.
• Use one quiet divider per source page and retain page references on selected claims; do not append a page number to every paragraph. Never expose parser names, hashes, item refs, confidence percentages, test notes, build status, or implementation commentary in report.md or the Notion page.
• Do not publish a report that fails validation. Spot-check the title, one heading transition, one formula, one embedded figure, one emphasized claim, and its page against the PDF. If extracted math is not valid LaTeX, use the cropped source formula instead of sending broken text to KaTeX.

Notion

When the user asks for Notion publishing, read references/notion.md. Fetch the real destination schema, deduplicate by DOI or title, write the report as the page body, and fetch the page afterward to verify it. Do not create or modify cloud content without the user's request.

When no connector is available, deliver report.md. Use the REST fallback only if the user explicitly chooses it and supplies credentials through local environment variables.

Access boundary

Accept local PDFs the user is authorized to access. For authenticated publisher pages, ask for the downloaded PDF. Never bypass paywalls, logins, or access controls, and never present public metadata or an abstract as a full-paper read.

Handoff

Lead with the reading document or confirmed Notion page. Mention one remaining manual action only when necessary; keep validation logs and internal artifacts out of the user-facing handoff unless the user asks for debugging details.