Document Processing · head-to-head
docling vs pdf
Pick docling when the document is input: it parses PDFs (born-digital or scanned), DOCX, PPTX, XLSX, HTML, images and audio into one structured representation, and chunks a corpus for retrieval. Pick the Anthropic pdf skill when the document is output, or when you need to act on a PDF in place — fill a form, watermark, encrypt, split, merge. Reading and writing are different jobs and these two split cleanly along that line.
Pick docling
You need to get content OUT of documents: extraction, format conversion, table recovery, OCR on scanned pages, or chunking a document set for RAG.
Pick pdf
You need to produce or modify a PDF: generate one, fill a form, watermark, encrypt, split, merge, or extract from an encrypted file.
Side by side
How they differ
| docling | ||
|---|---|---|
| Direction | Read and convert — documents in, structured Markdown or JSON out | Read and write — produces and manipulates real PDFs |
| Formats | PDF, DOCX, PPTX, XLSX, HTML, Markdown, AsciiDoc, CSV, images, audio, XML | PDF only |
| Scanned pages / OCR | Yes — handles scanned PDFs | Extraction works on text layers; OCR is not its job |
| RAG chunking | Built in | No |
| Form filling / encryption | No | Yes |
| Runs locally | Yes — Python package, no API key | Yes — bundled scripts |
| Provenance | Verified (Docling project, LF AI & Data) | Official (Anthropic) |
Direction
- docling
- Read and convert — documents in, structured Markdown or JSON out
- Read and write — produces and manipulates real PDFs
Formats
- docling
- PDF, DOCX, PPTX, XLSX, HTML, Markdown, AsciiDoc, CSV, images, audio, XML
- PDF only
Scanned pages / OCR
- docling
- Yes — handles scanned PDFs
- Extraction works on text layers; OCR is not its job
RAG chunking
- docling
- Built in
- No
Form filling / encryption
- docling
- No
- Yes
Runs locally
- docling
- Yes — Python package, no API key
- Yes — bundled scripts
Provenance
- docling
- Verified (Docling project, LF AI & Data)
- Official (Anthropic)
Open docling or pdf for the composite score, license, and exact install commands, or browse the full Document Processing category.
Verdict
The bottom line
Use both, in sequence, for the common real workflow: docling to read a messy input corpus into structured text, then the pdf skill to emit the deliverable. The mistake is reaching for pdf when the input is a scanned contract, or for docling when the deliverable is a filled form — each will do a poor job of the other half. If you can only install one, install the one matching whether documents are arriving or leaving.
FAQ
Common questions
Is docling better than the Anthropic pdf skill?
Neither is better; they point in opposite directions. docling is the strongest reader and converter available, covering ten-plus formats including scanned PDFs and RAG chunking. The Anthropic pdf skill is the one that writes: generating PDFs, filling forms, watermarking, encrypting. Pick by whether the document is your input or your output.
Can docling replace all four Anthropic document skills?
For reading, largely yes — it parses DOCX, PPTX and XLSX as well as PDF. For producing polished Office files it cannot: docx, xlsx and pptx generate real formatted documents with formulas, layouts and speaker notes, which docling does not do.
Does docling need an API key?
No. It runs locally as a Python package (3.10+), and the skill ships inside that package so it updates with it. There is an optional remote service client if you prefer to run conversion on a server.
Related
More head-to-head comparisons
- firecrawl vs browser-use
firecrawl for hosted, JavaScript-rendered extraction at scale; browser-use for driving a real browser through logins, forms, and multi-step flows.
- supabase-skill vs planetscale-skills
supabase-skill for Postgres with row-level security, auth, and realtime; planetscale-skills for MySQL with Git-style schema branching at scale.
- stripe-best-practices vs paddle-agent-skills
stripe-best-practices for maximum control over payment flows (you own tax); paddle-agent-skills for merchant-of-record billing (Paddle handles global sales tax and VAT).
- semgrep vs trail-of-bits
semgrep runs a real static-analysis engine and writes custom rules; trail-of-bits brings a security firm's review method to a diff. Findings versus judgement.
- playwright-cli vs Browserbase browser
playwright-cli for a local browser you fully control and an existing Playwright suite; Browserbase browser for hosted sessions with CAPTCHA solving and residential proxies.
- postgres vs supabase-skill
The postgres skill for engine-level decisions — types, indexes, partitioning, extensions; supabase-skill for the platform layer — row-level security, auth, realtime.
- drawio vs mermaid-diagrams
drawio produces an editable .drawio file someone can rearrange later; mermaid keeps the diagram as text inside the document it belongs to.
- Stripe Connect vs Paddle
Stripe Connect for control over multi-party payment flows you operate yourself; Paddle for handing global tax and compliance to a merchant of record.
Browse all best-skill guides or the full directory.