pdf-toolkit

Deterministic, structural PDF operations. Use this skill for programmatic work where you know exactly what you want done. For a natural-language rewrite, first draft the replacement content with ordinary reasoning, then use the explicit extract/generate/merge operations here to create the final PDF.

Use the inline Python examples with execute_code in a restricted channel. Keep files in the active workspace and call publish_artifact with the finished PDF. Shell commands below require an available exec_command; they are optional shortcuts, not a reason to request host execution. If the sandbox or a required library is unavailable, report the limitation rather than retrying outside the sandbox.

Decide the operation

Path A: Extract

python {baseDir}/scripts/extract.py /path/to/doc.pdf --json

Output:

{
  "pages": 12,
  "metadata": {"title": "...", "author": "..."},
  "text": [
    {"page": 1, "content": "..."},
    {"page": 2, "content": "..."}
  ],
  "tables": [
    {"page": 3, "rows": [["..."], ["..."]]}
  ]
}

Text uses pdfplumber (already in default dependencies) which preserves column layout better than naive PDF text extraction. Tables use pdfplumber.extract_tables() with default settings; for tricky layouts pass --tables-strategy lines|text|explicit to switch detection mode.

For OCR (scanned PDFs), this skill does not include Tesseract — use the sibling skill that wraps an OCR engine (out of scope here).

Path B: Merge / Split

Merge full files:

python {baseDir}/scripts/merge.py a.pdf b.pdf c.pdf --out combined.pdf

Or merge specific page ranges with the manifest form:

python {baseDir}/scripts/merge.py manifest.json --out combined.pdf

manifest.json:

[
  {"file": "a.pdf", "pages": "1-3"},
  {"file": "b.pdf", "pages": "5,7,9-11"},
  {"file": "c.pdf"}
]

Page ranges are 1-based, comma-separated, hyphen for ranges. Omit pages to include the whole file. Splits use the same syntax in reverse:

python {baseDir}/scripts/split.py input.pdf --pages "1-3,7,10-12" --out output_dir/

Each range writes one output file: output_dir/input_001.pdf, output_dir/input_002.pdf, …

Path C: Form fill

python {baseDir}/scripts/form_fill.py form.pdf data.json --out filled.pdf

data.json maps field name → string value:

{
  "applicant_name": "Wei E.",
  "submission_date": "2026-05-06",
  "agreed": "Yes"
}

The script discovers fields via pypdf.PdfReader.get_fields() and updates them with update_page_form_field_values(). Fields not present in the JSON are left untouched. Run with --list-fields to enumerate the form's fields without filling.

Caveats:

• /Btn checkbox fields take the export value (often Yes, On, or 1) rather than true — inspect with --list-fields to discover.
• AcroForm fills only. XFA forms (used by some legal templates) require Adobe-specific tooling and are out of scope.
• Some signed PDFs invalidate the signature when fields change. Strip signatures explicitly with --clear-signatures if that is intended.

Path D: Generate from scratch

Use reportlab directly when you need a new PDF:

from reportlab.pdfgen import canvas
from reportlab.lib.pagesizes import LETTER
from pathlib import Path

c = canvas.Canvas(str(Path("out.pdf")), pagesize=LETTER)
c.setFont("Helvetica-Bold", 18)
c.drawString(72, 720, "Q3 Review")
c.setFont("Helvetica", 11)
c.drawString(72, 696, "Revenue grew 18% year over year.")
c.showPage()
c.save()

For Chinese text, register a CJK font instead of Helvetica. ReportLab's CID font works without downloading fonts or reading a user font directory:

from reportlab.pdfbase import pdfmetrics
from reportlab.pdfbase.cidfonts import UnicodeCIDFont
from reportlab.pdfgen import canvas

pdfmetrics.registerFont(UnicodeCIDFont("STSong-Light"))
c = canvas.Canvas("report.pdf")
c.setFont("STSong-Light", 14)
c.drawString(72, 760, "季度报告：收入增长")
c.save()

CID fonts rely on PDF reader CJK support. When an embedded font is required, use a licensed font already available inside the workspace or permitted system font roots. Validate the text with pypdf before publishing.

For tables, headers/footers, and multi-column layouts, switch to reportlab.platypus (SimpleDocTemplate, Paragraph, Table, PageBreak). See references/reportlab.md.

Natural-language changes

This public entry deliberately keeps the PDF mutation step deterministic. For requests such as "make the title shorter", inspect the source page, draft the replacement text with the model, and generate a new document with the reviewed content. Do not claim that arbitrary in-place page rewriting is available.

Common pitfalls

Boundaries

• This skill works with text-based and form-based PDFs. Scanned image PDFs need OCR before any text path produces results.
• Encrypted PDFs are read-only here. Decryption requires the user-supplied password and is out of scope for this skill.
• For PDF-to-image rendering, use a separate skill that wraps Poppler or PyMuPDF.
• Digital signature operations (signing, verifying, revoking) are out of scope.