PDF Monster

Core Rule

Treat PDF files as source artifacts that must be converted into model-readable evidence before analysis. Do not create output/, analysis/, pages/, or similar folders in the user's working directory unless the user explicitly asks to save extracted artifacts.

Use the bundled scripts/analyze_pdf.py first. Resolve this path from the skill root, not from the user's current working directory. It prints JSON to stdout, extracts text in-place, and writes only image artifacts to an OS temporary directory unless --save-to is provided.

Quick Start

Run from this skill directory or replace scripts/analyze_pdf.py with its absolute path:

bash
python3 scripts/analyze_pdf.py path/to/file.pdf --json

When the agent is running from another directory, use the absolute installed path. In Claude Code, ${CLAUDE_SKILL_DIR} points at this skill directory:

bash
python3 ${CLAUDE_SKILL_DIR}/scripts/analyze_pdf.py path/to/file.pdf --json

Setup

Python 3 is required. Before first use, check whether PyMuPDF is already available:

bash
python3 -c "import fitz"

If that fails and the agent is allowed to run pip/network installs, install the recommended dependency from this skill's requirements.txt. In Claude Code:

bash
python3 -m pip install -r "${CLAUDE_SKILL_DIR}/requirements.txt"

In other agents, use the absolute path to this skill's requirements.txt.

If dependency installation is not allowed, continue with the script anyway; it can use Poppler fallbacks when available. Optional system tools improve coverage when PyMuPDF is unavailable or OCR is needed, but do not install system packages automatically:

• Poppler: pdfinfo, pdftotext, pdftoppm, pdfimages
• Tesseract: tesseract plus any needed language data, such as eng or kor

For visual-heavy or scanned documents:

bash
python3 scripts/analyze_pdf.py path/to/file.pdf --render-pages all --ocr auto --json

For Korean/English OCR, use Tesseract language data and pass:

bash
python3 scripts/analyze_pdf.py path/to/file.pdf --render-pages all --ocr auto --ocr-lang kor+eng --json

For text-only inspection with no temporary image artifacts:

bash
python3 scripts/analyze_pdf.py path/to/file.pdf --render-pages none --no-extract-images --ocr never --json

For slide decks or PDFs with repeated logos/icons, reduce embedded image noise while keeping page renders available:

bash
python3 scripts/analyze_pdf.py path/to/file.pdf --render-pages all --min-image-area 10000 --dedupe-images --json

For selected pages:

bash
python3 scripts/analyze_pdf.py path/to/file.pdf --pages 1,3-5 --render-pages all --json

Persist artifacts only when the user asks for reusable files:

bash
python3 scripts/analyze_pdf.py path/to/file.pdf --save-to ./pdf-monster-artifacts --json

Reading Workflow

1. Run analyze_pdf.py and capture stdout JSON.
2. Check warnings, page_count, artifact_root, pages_needing_visual_review, and each page's text_chars, ocr_text_chars, render_path, embedded_images, needs_visual_review, and visual_review_reasons.
3. Use text as the primary evidence when it is complete enough.
4. Inspect ocr_text when a page has little extracted text or appears scanned.
5. Open render_path for pages marked needs_visual_review, and for pages where layout, charts, handwriting, equations, tables, screenshots, or visual placement could change the answer.
6. Open embedded_images[].path when the PDF contains standalone figures that may be clearer than the page render.
7. Cite page numbers from the manifest when answering.
8. Delete temporary artifacts after finishing if artifact_policy is temporary.

The JSON includes a cleanup_command such as:

bash
rm -rf -- /tmp/pdf-monster-...

Run it only after the image paths are no longer needed. Never delete a --save-to directory unless the user asks.

Finalization Checklist

Before answering, complete these checks:

• If pages_needing_visual_review is non-empty, inspect the relevant render_path or rerun selected pages with --render-pages all before making claims that depend on layout, diagrams, tables, or figures.
• If rendered pages or extracted images were created with artifact_policy: temporary, run cleanup_command after the visual evidence is no longer needed.
• Cite page numbers for document claims. If the evidence is incomplete because OCR or rendering is unavailable, say that clearly.
• Do not expose names, student IDs, email addresses, signatures, or other personal data in summaries unless the user specifically asks for that information.
• Do not paste raw JSON, temporary artifact paths, or long extracted text into the final answer unless the user asks for those details.

Script Behavior

analyze_pdf.py prefers PyMuPDF when available. It falls back to Poppler CLI tools where possible:

• pdftotext for text extraction
• pdfinfo for page count
• pdftoppm for page renders
• pdfimages for embedded image extraction
• tesseract for OCR when installed

OCR is optional. If tesseract is missing, report the missing OCR capability and continue with text extraction and rendered page images.

Other Agent Use

Install this folder where the agent discovers skills, or reference the absolute path in custom instructions:

• Claude Code: ~/.claude/skills/pdf-monster or .claude/skills/pdf-monster
• Codex: $CODEX_HOME/skills/pdf-monster or ~/.codex/skills/pdf-monster
• Pi Coding Agent: ~/.pi/agent/skills/pdf-monster, ~/.agents/skills/pdf-monster, or a skills settings entry
• OpenClaw: ~/.openclaw/skills/pdf-monster, <workspace>/skills/pdf-monster, or skills.load.extraDirs
• Hermes Agent: ~/.hermes/skills/pdf-monster or skills.external_dirs

For agents without a native skill loader, give them this folder and a short instruction:

text
Use pdf-monster/SKILL.md. For PDF tasks, run:
python3 /absolute/path/to/pdf-monster/scripts/analyze_pdf.py <pdf> --json
Read the JSON from stdout. Use temporary image paths only while analyzing, then clean them up.
Do not create output folders in the user's working directory unless explicitly asked.

For OpenCode or similar agents, keep the repository checked out somewhere stable and reference the absolute path to SKILL.md or scripts/analyze_pdf.py in the agent's custom instructions. The only hard requirement is Python 3. PyMuPDF is the recommended dependency; Poppler and Tesseract improve fallback and OCR coverage.