paper2ppt Skill

Generate a professional academic literature review PowerPoint (English PPT) from a paper DOI, PMID, PMCID, URL, local PDF, or search query.

Language Policy

• PPT output language: English only. All slide titles, body text, labels, and citations are in English.
• Input language: Accepts Chinese or English input from the user.
• Paper language: Works with papers in any language. If the paper is in Chinese, translate extracted content to English for the PPT.
• Journal name: Preserve the original journal name (e.g., "中国循证医学杂志" stays as-is), prefix with English if known.

Robustness Rules (Read First)

Apply these to every phase.

1. Every external call must have a timeout. Use curl --connect-timeout 10 --max-time 30 for HTTP requests.
2. Every external call must have a fallback. If CrossRef fails, try NCBI. If NCBI fails, try WebSearch.
3. Never leave a slide blank. If figures are missing, add text content instead. If text is missing, use available metadata.
4. Never fabricate data. If IF is not found, show "IF: N/A". Never guess or infer IF.
5. Validate before running. Check file paths, check figure files are non-empty, syntax-check the script with node --check.
6. Kill long-running processes. Set timeout: 60000 on all curl/Node.js calls.
7. Handle special characters. Strip HTML tags, escape problematic characters in text.
8. Check output file. Verify the .pptx file exists and has non-zero size.

Workflow Overview

User provides paper source (DOI/PMID/PMCID/URL/PDF path/title)
        │
        ▼
Phase 0: Identify paper → cross-validate → confirm → resolve DOI
        │
        ▼
Phase 1: Fetch metadata (CrossRef + NCBI cross-validation)
        │
        ▼
Phase 2: WebSearch → Impact Factor + JCR quartile
        │
        ▼
Phase 3: Extract figures + tables + section text
        │   ├─ 3.1 Online: publisher HTML → figure URLs
        │   ├─ 3.2 PDF available: PyMuPDF embedded images
        │   ├─ 3.3 Tables: pdfplumber → addTable reconstruction
        │   └─ 3.4 Section text: WebFetch / PyMuPDF text extraction
        │
        ▼
Phase 4: Generate PPTX → plan → write script → run → QA → fix loop

Phase 0: Paper Identification (with Cross-Validation)

0.1 Input Type Detection

0.2 Local PDF Entry (NEW)

When the user provides a local PDF file path:

# Verify the file exists
ls -la "{PDF_PATH}"

Step 1 — Extract PDF metadata:

import fitz
doc = fitz.open("{PDF_PATH}")
meta = doc.metadata  # title, author, subject
first_page = doc[0].get_text()[:2000]  # first page text for DOI/title search

Step 2 — Search for DOI in first page text:

import re
doi_match = re.search(r'10\.\d{4,}/[^\s\)\]]+', first_page)
doi = doi_match.group(0) if doi_match else None

Step 3 — If DOI found: Proceed to Phase 0.3 (resolve and verify). Step 4 — If no DOI: Extract the paper title from the first page (usually the first large text block before "Abstract" or authors). Use it to search CrossRef:

curl -s "https://api.crossref.org/works?query.title={URL_ENCODED_TITLE}&rows=3"

Pick the best match by comparing titles. Then proceed to verification.

0.3 Resolution to DOI

0.4 Paper Verification (CRITICAL — Do Not Skip)

Never trust a single source without cross-validation. Wrong papers happen when identifiers are mismatched, search returns the wrong result, or metadata is stale.

Verification procedure:

1. Fetch title from CrossRef using the resolved DOI:

   curl -s "https://api.crossref.org/works/{DOI}?mailto=user@example.com" | python -c "import sys,json; d=json.load(sys.stdin); print(d['message']['title'][0])"
   

2. Fetch title from NCBI (if PMID or PMCID was available):

   curl -s "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esummary.fcgi?db=pubmed&term={DOI}&retmode=json" | python -c "import sys,json; d=json.load(sys.stdin); print(d['result'][list(d['result'].keys())[0]]['title'])"
   

3. Compare titles — if both sources are available, check they refer to the same paper:

   • Extract ~50 chars from each title
   • Compare normalized strings (lowercase, remove punctuation)
   • If match confidence < 80%, ask the user to confirm: "I found two candidates: ... Which one is correct?"
   • If only one source is available, show the title to the user and ask: "Is this the correct paper?"

4. User confirmation: After verification (successful or single-source), show the user:

   Paper: {Title}
   Authors: {Authors}
   Journal: {Journal} ({Year})
   DOI: {DOI}
   Proceed? (Y/N)
   

   Wait for confirmation. If the user says no, stop and ask for a corrected identifier.

5. If the user provides a PDF: Extract the title from the first page and the DOI. Cross-validate them with CrossRef. If they mismatch, warn the user.

CRITICAL RULE: If you are uncertain about the paper identity (>80% confident), do not proceed. Ask the user.

Phase 1: Metadata Fetching

1.1 Primary: CrossRef API

curl -s --connect-timeout 10 --max-time 30 \
  "https://api.crossref.org/works/{URL_ENCODED_DOI}?mailto=user@example.com"

Fields to extract:

Strip HTML tags from abstract: Use sed 's/<[^>]*>//g' or JavaScript .replace(/<[^>]*>/g, '').

Author formatting:

• List first 3 authors, then "et al." if more than 3
• Format: "Given Family", e.g. "G. Kucsko, J. Smith, et al."
• If author names are in Chinese characters, use the romanized form from CrossRef.

1.2 Fallback: NCBI E-utilities

If CrossRef is unreachable:

curl -s "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esummary.fcgi?db=pubmed&term={DOI}&retmode=json"

Extract: title, authors, journal, date, volume, issue, pages, PMID.

1.3 Last Resort: WebSearch

If both CrossRef and NCBI fail, use WebSearch with the DOI or paper title to find the paper info from publisher snippets. Build the best-effort title slide and mark it as "metadata from web search."

Phase 2: Impact Factor and JCR Quartile Lookup

Step 1: Prepare search query

"[journal name]" "impact factor" {year}

Step 2: Extract IF

Parse snippets for "IF: