Extract Pdf

You are the PDF Extractor. Your job is to turn any PDF into a structured workspace that the calling agent (Claude) can read directly — markdown files for the textual content (with tables, headings, and lists preserved), PNG files for the visual content. There is no separate vision API in the loop: Claude itself reads the rendered page images with its built-in Read tool.

Why this skill exists

Naive PDF text extraction loses figures, tables, formulas, layout, and anything in scanned pages. The classic workaround is to send each page image to an external vision model, which adds API keys, cost, and rate limiting. Heavy ML pipelines (Marker, Nougat, MinerU, Docling) need GBs of model weights and a GPU. Claude is already multimodal, so the right move is to compose lightweight tools and let Claude handle anything that needs vision:

1. Render each page to PNG locally with PyMuPDF. No system dependencies, no poppler, no ML.
2. Extract structured Markdown with pymupdf4llm — the official LLM-targeted PDF→Markdown extractor from the PyMuPDF maintainers. It detects tables, headings, lists, and code blocks and preserves them as proper Markdown so Claude can read them at a glance instead of squinting at column-aligned spaces.
3. Run pdfminer.six in parallel on the same page set as a complementary plain-text extractor. PyMuPDF and pdfminer use different layout-analysis algorithms, and when one mangles a page (ligatures, columns, vertical text, weird encodings) the other often nails it. Disagreement between the two is itself a useful signal.
4. Hand all of it to Claude as files it can Read on demand. For math, scanned pages, or figures where text extraction can't be exact, the agent reads the page image directly — no extra OCR step.

This makes PDF analysis a one-command operation that any future skill or task can reuse.

When to use

Trigger this skill whenever a task involves a PDF and any of:

• "summarize / read / analyze / explain this PDF"
• "extract text / tables / figures from a PDF"
• "what does the chart on page 4 show?"
• "convert this PDF to something I can work with / to markdown"
• the PDF has scanned pages, math, diagrams, or complex layout
• another skill (e.g. index-note for paper/book) needs structured PDF content

Skip it only when the user already has the content extracted, or when the PDF is just a vehicle for a tiny text snippet and there is no visual content to worry about.

Output layout

Every run produces this directory:

/
├── manifest.json            # index of everything; ALWAYS read this first
├── full_text.md             # pymupdf4llm full-document Markdown
├── full_text_pdfminer.md    # pdfminer.six plain-text complement (saved as .md)
├── pages/
│   ├── page_001.md          # per-page Markdown (aligned with image by number)
│   ├── page_002.md
│   └── ...
└── images/
    ├── page_001.png         # full-page render of page 1
    ├── page_002.png
    └── ...

manifest.json is the entry point. Key fields:

• source_pdf, filename, total_pages_in_pdf, processed_pages
• pdf_metadata — title/author/subject/keywords/creator/producer
• full_text_path — primary Markdown file (pymupdf4llm)
• full_text_format — "markdown" if pymupdf4llm succeeded, "plain" if it had to fall back to bare PyMuPDF text (still in .md files, just without table/heading detection)
• full_text_pdfminer_path — pdfminer complement (null if pdfminer was unavailable or --no-pdfminer was passed)
• full_text_pdfminer_chars — character count of the pdfminer output, useful for quick comparison
• tables_detected — total number of tables pymupdf4llm found across all processed pages
• pages[] — for each processed page: page number, text_path, text_chars, table_count (when > 0), image_path, image_bytes
• options — what was actually extracted (DPI, text/images flags, whether pymupdf4llm and pdfminer ran)

Three surfaces, when to use which

| File | Source | Strengths | |---|---|---| | pages/page_NNN.md + full_text.md | pymupdf4llm | Real Markdown with tables, headings, lists preserved. Aligned with images/page_NNN.png by page number. Start here. | | full_text_pdfminer.md | pdfminer.six | Plain text from a different extractor. Cross-check when something looks wrong in the primary extraction — the two often disagree on tricky layouts (ligatures, multi-column, vertical text), and the disagreement itself flags pages worth looking at visually. | | images/page_NNN.png | PyMuPDF render | The ground truth. Read directly for: math equations, figures, tables that look mangled in either text file, scanned pages, anything where layout matters. Claude reads PNGs natively — no separate OCR or vision API. |

Workflow

Step 1 — Plan the extraction

Before running anything, decide:

1. Source: a local path (/abs/path/file.pdf) or a URL (https://...). Both are accepted directly — no manual download needed.
2. Output directory: pick a stable workspace, e.g. ./.pdf_extracts/{slug}/. Reuse the same directory across runs of the same PDF so prior extractions stay cached. If the user has not specified one, default to ./.pdf_extracts/{pdf_basename_without_ext}/.
3. Page range: if the PDF is large (>30 pages) and the user only cares about a section, pass --pages to avoid wasted work. Examples: --pages 1-5, --pages 1,3,7-10. Default is all pages.
4. Modality: by default extract both text and images. Use --no-images if the user explicitly wants text only (much faster on long PDFs), or --no-text if they only want page renders.

Step 2 — Run the extractor

The script must be invoked through uv so its dependencies are available without polluting the system Python. pymupdf4llm brings pymupdf along automatically, so two --with flags cover everything:

uv run --with pymupdf4llm --with pdfminer.six python ./skills/extract-pdf/scripts/extract_pdf.py \
  --input "" \
  --output-dir ""

Common variants:

# Pages 1-5, text-only, very fast
uv run --with pymupdf4llm --with pdfminer.six python ./skills/extract-pdf/scripts/extract_pdf.py \
  --input report.pdf --output-dir ./.pdf_extracts/report --pages 1-5 --no-images

# Hi-res rendering for fine print, tiny tables, or dense math
uv run --with pymupdf4llm --with pdfminer.six python ./skills/extract-pdf/scripts/extract_pdf.py \
  --input scan.pdf --output-dir ./.pdf_extracts/scan --dpi 220 --max-image-edge 3000

# URL input — downloaded automatically
uv run --with pymupdf4llm --with pdfminer.six python ./skills/extract-pdf/scripts/extract_pdf.py \
  --input "https://arxiv.org/pdf/2401.12345" --output-dir ./.pdf_extracts/2401.12345

The script prints the full manifest as JSON on stdout when it finishes. You can either parse that or read /manifest.json.

Step 3 — Read the manifest

Read /manifest.json

Check total_pages_in_pdf, processed_pages, tables_detected, and per-page text_chars. Pages with very small text_chars are likely scanned/image-only — for those you must read the image, not the text. full_text_format tells you whether the primary extraction is real Markdown ("markdown") or fell back to plain text ("plain").

Step 4 — Read the content (this is where Claude does the work)

Three surfaces, used together based on what the task needs.

When the task is textual

Start with the structured Markdown. Tables and headings are preserved as Markdown so they're scannable at a glance:

Read /full_text.md

For surgical reads of one page only, use the per-page file (aligned with the image of the same page):

Read /pages/page_004.md

If a passage looks wrong (gibberish, missing words, broken table, suspicious gaps), cross-check the same passage in the pdfminer complement:

Read /full_text_pdfminer.md

If the two disagree, that page's layout is tricky — go to the image to settle it.

When the task involves visuals

The Read tool natively supports PNG. To analyze a figure, table, chart, equation, or scanned page, just read the page image:

Read /images/page_004.png

You will see the page as an image and can describe, transcribe, or reason about its visual content directly. No external vision API is involved.

This is also the right move for:

• Math equations — text extractors can't reliably preserve LaTeX. Read the image and transcribe to $...$ or $$...$$ yourself.
• Complex tables — if the markdown table looks misaligned, the image is authoritative.
• Figures — describe what you see directly from the render.

When you need both

For tasks like "explain Figure 3 and quote the caption", read the markdown for the words and the image for the visual content, then synthesize. Cross-reference by page number — pages/page_NNN.md and images/page_NNN.png are aligned.

Step 5 — Be efficient about which pages to read

You do not need to read every page image. Be selective:

1. Start by reading full_text.md (or skim manifest.json to see page counts and table counts).
2. Identify which pages actually need visual inspection — usually those with:

• Suspiciously low text_chars (scanned or image-only)
• References to "Figure N", "Table N", "Eq. N" in the text
• Math content the user asked about
• Pages the user explicitly named

1. Read only those page images.

This keeps token usage proportional to the task, not to PDF size.

Examples

Example 1 — Quick summary of a research paper

uv run --with pymupdf4llm --with pdfminer.six python ./skills/extract-pdf/scripts/extract_pdf.py \
  --input "https://arxiv.org/pdf/2505.00949" \
  --output-dir ./.pdf_extracts/2505.00949

Then:

1. Read ./.pdf_extracts/2505.00949/manifest.json — see it has 12 pages, 3 tables detected
2. Read ./.pdf_extracts/2505.00949/full_text.md — get the prose with tables already in Markdown
3. Find references to "Figure 2" on page 4 → Read ./.pdf_extracts/2505.00949/images/page_004.png to describe it
4. Write the summary

Example 2 — Scanned contract (no embedded text)

uv run --with pymupdf4llm --with pdfminer.six python ./skills/extract-pdf/scripts/extract_pdf.py \
  --input ./contract_scan.pdf --output-dir ./.pdf_extracts/contract --dpi 200

The manifest will show text_chars: 0 (or near zero) for every page. Skip the text files and read each images/page_NNN.png directly — Claude will OCR by sight.

Example 3 — Just the table on page 7

uv run --with pymupdf4llm --with pdfminer.six python ./skills/extract-pdf/scripts/extract_pdf.py \
  --input report.pdf --output-dir ./.pdf_extracts/report --pages 7

Then Read ./.pdf_extracts/report/pages/page_007.md — pymupdf4llm will already have the table as a Markdown table. If it looks wrong, fall back to Read ./.pdf_extracts/report/images/page_007.png.

Example 4 — Long book, only the introduction

uv run --with pymupdf4llm --with pdfminer.six python ./skills/extract-pdf/scripts/extract_pdf.py \
  --input book.pdf --output-dir ./.pdf_extracts/book --pages 1-25 --no-images

Text-only is much faster and cheaper when no figures matter.

Options reference

| Flag | Default | Notes | |---|---|---| | --input | required | Local path or http(s) URL. URLs download automatically. | | --output-dir | required | Directory to write manifest.json, full_text.md, pages/, images/. | | --pages | all | Selection like 1-5,9,12-15 (1-indexed, inclusive). | | --dpi | 150 | Render DPI for page images. Bump to 200–300 for fine print, dense math, or small tables. | | --max-image-edge | 2000 | Caps the longest edge in pixels so individual page images stay readable but bounded. Set to 0 to disable. | | --no-images | off | Skip rendering — text only. Much faster for large PDFs. | | --no-text | off | Skip text extraction — images only. | | --no-pdfminer | off | Skip the complementary pdfminer.six pass. Useful if pdfminer hangs on a pathological PDF. | | --table-strategy | lines_strict | pymupdf4llm table detection. lines_strict (default) finds bordered tables reliably and never false-positives on prose; borderless table content still comes through as plain-text rows. lines is slightly looser. text is aggressive — only use it on table-heavy docs with clean columns; it can wrap entire prose pages in fake tables. |

Failure modes and fixes

• PyMuPDF not found → you forgot uv run --with pymupdf4llm --with pdfminer.six. Always invoke through uv exactly as shown.
• full_text_format: "plain" in the manifest → pymupdf4llm failed for this PDF (rare; check stderr) and the script fell back to bare PyMuPDF text. The per-page .md files still contain the text, just without table/heading structure. The PDF is still fully usable; you'll need to look at images to recover table layout.
• full_text_pdfminer.md is missing → pdfminer.six wasn't installed (add --with pdfminer.six), it raised on a malformed PDF (check stderr), or --no-pdfminer was passed. Primary extraction is unaffected.
• text_chars is 0 for every page → the PDF is a scan. Read the page images instead of the text files; Claude will OCR by sight.
• Page image is too small to read fine print → rerun with --dpi 220 (or higher) and a larger --max-image-edge.
• PDF is huge and the run is slow / costs a lot of tokens → narrow with --pages, or do a text-only first pass with --no-images to find the relevant pages, then re-extract just those pages with images.
• URL download fails → some publishers require auth or block User-Agents. Download manually and pass the local path.
• A borderless table you care about wasn't detected as a markdown table → the row data is still in full_text.md as plain-text lines (no information loss), and you can read images/page_NNN.png to see the layout. Only escalate to --table-strategy text if the document is overwhelmingly tables with clean columns — on prose-heavy PDFs that strategy hallucinates fake tables around entire paragraphs.

Reusability promise

Once this skill is in the project, no other skill or task should write its own pdfminer / pdf2image / PyMuPDF / pdfplumber boilerplate. If a workflow needs PDF content, it should call extract_pdf.py and read the resulting files. That is the whole point of the skill.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: KaihuaTang
• Source: KaihuaTang/Index.skill
• License: MIT

Install and usage instructions live in the source repository linked above.