Paper2html

paper2html — Paper to single-page project homepage (you-led, coordinated)

Turn a paper PDF into a self-contained, publish-ready single-page project website — the kind of paper homepage researchers commonly build on GitHub Pages. You are the lead author: this file is the recipe, not a fully automated script — there is no main.py, no renderer. The mechanical steps (parse / extract / QA) call the small tools under scripts/; understanding the paper, designing the page, and writing index.html are done by you (read the material and figures with Read, write index.html with Write), confirming with the user at key points via AskUserQuestion.

PDF
 → parse + extract        (parse_pdf.py: MinerU → clean.md + manifest.json + images/, gate 1)
 → you understand it + set the design   (read manifest + clean.md + look at figures) → pick a design language   [confirm design direction]
 → you hand-author the page   (per the references/ design-language and authoring rules) → index.html            [optional confirm]
 → QA validation          (validate.py: missing figures/broken links/content fidelity, gate 2) → fix per report, loop
 → single-page project homepage index.html (+ images/, deployable as-is)

How you run this skill

1. Step by step: run scripts for the mechanical steps via Bash (absolute paths, no cd needed); do the design and authoring yourself with Read/Write.
2. Compute WORKDIR inline at the top of each Bash block (each Bash call is a separate shell and does not share variables):

   WORKDIR="$(dirname "$pdf_path")/.paper2anything/html/$(basename "${pdf_path%.*}")"

   $pdf_path is the paper PDF the user gave (reset it in each block). The scripts live in ${SKILL_DIR}/scripts — SKILL_DIR is this skill's directory (see "Base directory for this skill: …" injected at the top of this skill); each Bash block is a separate shell, so in the blocks that use it export SKILL_DIR=<that directory> once at the top (set it inline each time, like WORKDIR).
3. Pause at decision points with AskUserQuestion: after understanding the paper, confirm the design direction (design language / primary color / emphasis); after the draft is done, you may confirm again.
4. Faithful to the manifest, you fill the gaps: use only the real material in manifest.json, don't fabricate numbers/authors/links; fields the manifest left empty (authors/abstract/links etc.) you complete from the full text of clean.md — you are the lead author, deterministic extraction is only scaffolding.

Step 0: Environment and credentials

Unified environment: all python commands run in paper2anything's unified conda environment (top-level environment.yml), prefixed with conda run -n paper2anything --no-capture-output.

Credentials are centralized in the package-root .env (copied from .env.example, already gitignored); export it once per new shell:

set -a; source <paper2anything package root>/.env; set +a

This skill only needs MINERU_API_TOKEN (to parse the PDF). Page design and authoring are done by you, calling no LLM API, so no OPENAI/LLM key is needed.

Dependency self-check:

conda run -n paper2anything --no-capture-output python -c "import requests, rich, dotenv, playwright, PIL" 2>&1
# render_check.py (Step 4 render self-check) needs the chromium engine; install it once before the first run:
#   conda run -n paper2anything --no-capture-output python -m playwright install chromium

Step 1: Parse + deterministic extraction (script, gate 1)

pdf_path="/path/to/paper.pdf"          # ← the user's paper PDF
WORKDIR="$(dirname "$pdf_path")/.paper2anything/html/$(basename "${pdf_path%.*}")"
conda run -n paper2anything --no-capture-output \
  python "${SKILL_DIR}/scripts/parse_pdf.py" "$pdf_path" --workdir "$WORKDIR"

Outputs (under $WORKDIR):

• clean.md — normalized full-text markdown (for you to read through)
• manifest.json — deterministically extracted facts: title/authors/affiliations/abstract/links/claims/figures/tables/ method_components/bibtex (appendix filtered; fields that couldn't be extracted are left empty for you to fill)
• images/ — the figure files the page references (figures + result-table screenshots), referenced by you as images/<name>
• parsed/ (MinerU raw parse, includes full.md for reuse on re-run), logs/

Optional: when you know the paper's canonical link, add --paper-url <URL> (no arxiv assumption; if omitted, links.paper stays empty); --code-url likewise.

After parsing, Read manifest.json and clean.md to read the full text.

Step 2: Understand the paper + set the design direction (you do this) [confirm]

1. Read manifest.json (verified material) + clean.md (full text); Read the key figures under images/ and judge with your own eyes which one works as the hero, which suit inlining, and which are result-table screenshots.
2. Read references/design-languages.md and establish a design concept for this paper (pick one design language or blend: magazine/product page/ terminal/poster/minimal/dashboard; set the primary color, structure, what element leads). Different papers should look different — don't reuse the previous one's style.
3. Use AskUserQuestion to confirm the design direction with the user (design language / primary color / what to emphasize). Author the page with the confirmation in hand.

Fill the gaps: if the manifest's authors/abstract/links are empty, complete them yourself from the full text of clean.md (this is your backstop).

Step 3: Hand-author index.html (you do this) [optional confirm]

Per the confirmed design direction, author $WORKDIR/index.html yourself with Write — a self-contained, deployable single-page website. First read references/html-authoring.md (hard constraints and pitfalls). Key points:

• Self-contained: reference images with the relative path images/<filename> (from the manifest's figures[].file / tables[].image, which parse_pdf already copied into images/); inline CSS or use a CDN; every <img> has a non-empty alt; leave no href="#".
• Light first screen (title/authors/affiliations/resource buttons), the main figure as a one-time large teaser, then abstract → claims → method → results → supporting figures → BibTeX (adjust to the paper's character, not mandatory).
• For result tables, prefer screenshots (tables[].image).
• figure CSS: don't let the border frame whitespace, and never stretch the image to fill whitespace (details in references/html-authoring.md).
• Faithful to the manifest, no fabrication; fill gaps from the full text.

After writing, you may use AskUserQuestion to show the user the design and structure (optional), and edit index.html directly per their feedback.