PDF Content Extraction Skill

Extract PDF content to clean, organized markdown.

Workflow

1. Extract — Run script to get raw content + metadata
2. Analyse — Review for patterns and issues
3. Clean — Manually remove noise (footers, watermarks, branding)
4. Organise — Restructure fragmented content
5. Output — Deliver clean markdown

Note: Only Step 1 uses a script. Steps 2–5 are performed manually by Claude reading and rewriting content. Do not write cleanup scripts.

Step 1: Extract

python /mnt/skills/user/pdf-extract/scripts/extract_pdf.py \
    /mnt/user-data/uploads/{filename}.pdf \
    /home/claude/extracted/

Options:

Option	Description
--pages 1-10	Extract specific page range
--method pymupdf4llm	Force primary extractor (better formatting)
--method pymupdf	Force fallback (more reliable for scanned PDFs)
--min-image-size 100	Skip images smaller than 100px (filters icons)

Output:

/home/claude/extracted/
├── {filename}.md      # Raw markdown with YAML frontmatter
├── metadata.json      # Structured metadata
└── images/            # Extracted images (if any)

Step 2: Analyse

Read the extracted markdown:

cat /home/claude/extracted/{filename}.md

Check YAML frontmatter for:

• extraction_method — Which extractor was used
• total_pages — Document length
• has_outline — Bookmarks exist (helps with structure)
• total_images — Number of images

Identify issues requiring cleanup:

• Repeated footers/headers on every page
• Watermarks, branding, page numbers
• Fragmented sentences across line breaks
• Malformed tables
• Image markers needing repositioning

Step 3: Clean

IMPORTANT: Manual cleanup only.

• Do NOT write Python scripts to clean the content
• Do NOT use sed, awk, or regex replacement commands
• Do NOT copy-paste the raw content and run substitutions

Instead: Read the extracted content, understand it, then write a clean version from scratch, omitting the noise as you write.

Why manual? Each PDF has unique patterns. Claude makes better contextual decisions than automated rules — knowing what's noise vs. legitimate content, handling edge cases, and preserving meaning.

Process:

1. Read through the extracted markdown completely
2. Identify repeated noise (footers, headers, branding, page numbers)
3. Note the actual content structure (sections, flow, key information)
4. Write the clean output directly, skipping noise as you go

Load references as needed for pattern recognition:

Repeated elements & source-specific patterns: See cleanup-patterns.md

• Use when: footers, headers, SME/PMT branding detected

Text fragmentation: See sentence-reflow.md

• Use when: sentences split across lines or pages

Table issues: See table-formatting.md

• Use when: tables have missing delimiters, broken structure

Image handling: See image-handling.md

• Use when: document contains images to process

Step 4: Organise

While writing the clean output, apply these formatting principles:

Heading Hierarchy

• Use # → ## → ### consistently
• Don't skip levels
• Remove redundant numbering if using markdown headers

Paragraph Flow

• Single blank line between paragraphs
• Remove orphan lines (single words alone)
• Merge related short paragraphs

Image Placement

Convert markers to proper markdown:

<!-- Before -->
<!-- IMAGE: images/page003_img001.png (450x280px) -->

<!-- After -->
![Figure 1: Description](./images/page003_img001.png)

View each image with view tool to write accurate alt text.

Step 5: Output

Write Clean File

After reading and mentally processing the extracted content, write the clean markdown directly to a file:

# Write clean content to file (Claude creates this content)
cat > /mnt/user-data/outputs/{filename}_clean.md << 'EOF'
# Document Title

[Clean content goes here - written by Claude, not copied]

EOF

Or use the create_file tool to write the clean content directly.

Copy Images (if applicable)

mkdir -p /mnt/user-data/outputs/images/
cp -r /home/claude/extracted/images/* /mnt/user-data/outputs/images/

Quality Check

• No repeated footers/headers
• No standalone page numbers
• No watermarks or branding
• Sentences properly rejoined
• Tables intact and readable
• Images converted to markdown syntax
• Heading hierarchy logical

Summary to User

Include:

• Pages extracted
• What was cleaned (types of noise removed)
• Images included (remind about images/ folder requirement)
• Any limitations noted

Error Handling

Error	Cause	Solution
"File not found"	Wrong path	Check /mnt/user-data/uploads/
"Invalid PDF header"	Not a PDF	Inform user file is invalid
"Extraction failed"	Protected/corrupted	Try --method pymupdf
Empty output	Scanned PDF	Inform user, suggest OCR

Special Cases

Scanned/Image PDFs

If extraction_method shows pymupdf (fallback) with minimal text:

• PDF is likely scanned/image-based
• Inform user OCR tools may be needed

Large Documents (50+ pages)

Consider extracting in ranges:

python extract_pdf.py doc.pdf ./out1/ --pages 1-25
python extract_pdf.py doc.pdf ./out2/ --pages 26-50

Multi-Column Layouts

Verify reading order makes sense. pymupdf4llm handles columns reasonably but may interleave incorrectly.

Output Format

Final markdown structure:

# {Document Title}

## {First Section}

{Clean content...}

## {Second Section}

{Clean content...}

---

*Source: {filename}.pdf | Extracted: {date}*