pdf-to-expert-report

name

pdf-to-expert-report

description

Transforms a research paper PDF into a comprehensive, visually-rich expert report by extracting ALL content (text, figures, diagrams, tables, equations, algorithms) using parallel vision agents, then synthesizing into a structured report optimized for PhD-level readers who lack domain expertise but are eager to master the paper in one reading. Trigger this skill whenever a user: - Provides any .pdf file path and wants to understand, analyze, summarize, or report on it - Says "read this paper", "explain this PDF", "brief me on this research", "create a report from this paper", "help me understand this", "master this paper", "extract knowledge from" - Wants deep understanding of academic or technical content from a PDF - Asks "what does this paper say about X" or "what are the key findings in this PDF" - Provides a PDF and asks ANY analytical or comprehension question about its content Always invoke this skill proactively — even for "quick summaries", the parallel extraction approach catches visual content (diagrams, charts) that plain text extraction misses. This skill is the right choice any time a PDF is involved and understanding is the goal.

PDF to Expert Report

A lossless, visually-aware pipeline that turns any research paper PDF into a report designed to let a PhD reader in ANY field master the paper in one sitting. Built on the insight that research papers communicate critical information through diagrams, charts, and figures — not just text — this skill uses vision-capable agents to see the full rendered page.

---

Overview of the Pipeline

PDF
 │
 ├─ Step 0: Render pages to PNG (PyMuPDF, 120 DPI)
 │          → preserves all visual content
 │
 ├─ Step 1: Read first 4 pages → identify paper, structure, scope
 │
 ├─ Step 2: Launch N parallel extraction agents (one per ~20 pages)
 │          → each agent sees and describes ALL content on their pages
 │          → each agent writes extract_pages_NNN_MMM.md
 │
 └─ Step 3: Read all extractions → write comprehensive report
            REPORT_<title>.md

---

Step 0: Setup and Rendering

Check dependencies

python -c "import fitz; print('PyMuPDF ok')" 2>&1
# If missing: pip install PyMuPDF

Render all pages to PNG

Run the bundled render script:

python "<skill_dir>/scripts/render_pdf.py" --input "<pdf_path>" --output "<pdf_dir>/pdf_pages/"

This creates numbered files: page_001.png, page_002.png, … in pdf_pages/. Use --dpi 150 if the paper has small text, dense tables, or intricate diagrams. The script also prints the total page count — use this to plan agent batches.

---

Step 1: Orient with First Pages

Read page_001.png through page_004.png using the Read tool. Note:

• Full title, authors, institution, venue/year
• Abstract — the paper's own one-paragraph summary of contributions
• Table of contents / section structure (if present)

This briefing is essential for writing better agent prompts in Step 2.

---

Step 2: Parallel Extraction Agents

Batch size

Pages	Agents	Batch size
1–30	2	~15 pages
31–60	3	~20 pages
61–100	5	~20 pages
101–200	8–10	~20 pages

Output files

Each agent writes to: <pdf_dir>/extract_pages_NNN_MMM.md

Agent prompt template

Read references/extraction-prompt.md for the full template. Substitute into the template:

• {PAPER_TITLE} — full title from Step 1
• {ABSTRACT_SNIPPET} — first 2–3 sentences of abstract
• {START_PAGE}, {END_PAGE} — agent's page range
• {PNG_FILE_LIST} — full absolute paths, one per line
• {OUTPUT_FILE} — absolute path to extraction output file

Launch ALL agents in a single message

Use multiple Agent tool calls in one response so they run truly in parallel. Do NOT launch agents one at a time — the whole point is parallel execution.

While agents run

• Sketch the report skeleton (Background, Problem sections) using the abstract
• Check references/report-guide.md to internalize the full report structure

---

Step 3: Synthesize the Report

Read all extract_pages_*.md files, then write the full report in one pass.

Output path

<pdf_dir>/REPORT_<slug>.md where slug = paper title lowercased, spaces → underscores

Report requirements — non-negotiable

These are the properties that make the report useful to a domain-outsider PhD:

1. Generous background — build the reader's mental model from scratch. Assume PhD intelligence but zero domain vocabulary. Use analogies to familiar concepts.

2. Visual structure — no section longer than 3 paragraphs without a list, table, ASCII diagram, or callout box. Walls of text fail the reader.

3. Every figure interpreted — do not just mention "Figure 3 shows the architecture." Describe what is in the figure AND what insight it conveys AND what claim it supports.

4. Numbers in context — "49% accuracy" means nothing. Say "49% — vs. 34% prior best, meaning this system correctly handles 45% more cases than the previous leader."

5. Claim-evidence pairing — every major claim must be followed by the evidence (experimental result, proof, ablation) that supports it.

6. Actionable implications — end with concrete takeaways: what should a practitioner, researcher, or decision-maker DO differently after reading this paper?

7. Honest limitations — what does the paper NOT show? What assumptions could break?

Follow references/report-guide.md for the exact section template.

---

Final Checklist

Before reporting done:

• All figures described and interpreted (not just mentioned)
• All tables reproduced in markdown
• Background builds from first principles — no assumed domain knowledge
• Every key claim has supporting evidence cited from the paper
• Numbers are contextualized with baselines and plain-language meaning
• Glossary covers all introduced terminology
• Actionable Implications section is concrete and specific
• Report file saved; path reported to user

---

References

• references/extraction-prompt.md — Full extraction agent prompt (copy and customize)
• references/report-guide.md — Complete report section template with guidance
• scripts/render_pdf.py — Cross-platform PDF-to-PNG renderer