Claude's Document Skills: Excel, PowerPoint, Word, and PDF

The most-used Claude skills are ones most people don't know are skills: ask claude.ai for a spreadsheet, a slide deck, a formatted Word document, or a filled PDF, and what answers isn't the base model winging it — it's one of four Anthropic-built document skills (xlsx, pptx, docx, pdf) loading and driving code execution to produce a real file. They're also the clearest demonstration of what the skills format is for.

What they actually do

Each document skill is a normal SKILL.md-plus-resources package that encodes format craft: how a workbook should be structured, what makes a slide layout hold up, how document styles and sections work. When one activates, Claude writes and runs code against the Python document stack in the execution sandbox — openpyxl for Excel, python-pptx for PowerPoint, python-docx for Word, PDF tooling for forms and extraction — following the skill's procedure instead of improvising.

The difference is visible in the output. Without the skill, "make me a budget spreadsheet" gets you a Markdown table. With it, you get an .xlsx with separate input and summary sheets, real formulas, and formatting — a file you open in Excel and keep using.

Where they're available

On the API, a request looks like this (both betas required):

response = client.beta.messages.create(
    model="claude-opus-4-8",
    max_tokens=16000,
    betas=["code-execution-2025-08-25", "skills-2025-10-02"],
    container={"skills": [{"type": "anthropic", "skill_id": "pptx", "version": "latest"}]},
    tools=[{"type": "code_execution_20260521", "name": "code_execution"}],
    messages=[{"role": "user", "content": "Create a 5-slide QBR deck from these notes: ..."}],
)

The generated file lives in the container; the response carries its file ID, and you download it via the Files API. The full API mechanics — /v1/skills, limits, the no-network constraint — are in Skills on claude.ai and the API.

Getting good documents, not just documents

The skills handle format mechanics; content quality is still yours to specify. The patterns that consistently raise output quality:

1. Provide the skeleton. A slide-by-slide brief, a column list with example rows, or the section outline. The skill turns structure into format; it can't invent your argument.
2. Anchor to an existing file. Upload last quarter's deck or your report template and ask Claude to match it — the skills read existing documents with the same libraries they write with.
3. State the format rules once. Brand colors, font, "no more than four bullets per slide", "currency to two decimals" — explicit constraints are followed reliably.
4. Ask for a verification pass. "Before finishing, re-open the file and check every formula resolves and no cell shows an error" — the skill runs code, so it can actually check its own artifact.
5. Iterate on the file, not from scratch. Follow-ups ("make sheet 2 a pivot by region") edit the existing artifact rather than regenerating.

Limits worth knowing

The execution container is sandboxed: no network access, so a document skill can't pull live data mid-build — provide the data in the conversation or as an uploaded file. On the API, uploads run through the Files API and requests carry at most 8 skills. And very large or intricate documents (hundred-tab workbooks, pixel-perfect design decks) still hit the ceiling of what code-built documents do well — the skills raise the floor dramatically; they don't replace a designer.

Why they matter beyond documents

The document skills are the reference implementation for the whole format. If you're writing your own skill, read the xlsx SKILL.md in anthropics/skills first: one job, a tight procedure, deterministic work pushed into code, and boundaries stated plainly. Our annotated examples walk the same patterns at smaller scale.