Best agent skills for PowerPoint, Word, Excel and PDF, and which you may use outside Claude

Anthropic's four document skills, pptx, docx, xlsx and pdf, are the most complete agent skills for PowerPoint, Word, Excel and PDF files, but their license allows use only inside Anthropic's services: Claude.ai, Claude Code and the Claude API. If you work in Codex, Cursor, Gemini CLI or another agent, OpenAI's pdf skill is the one we list that you may copy and change (Apache-2.0), and it covers PDF only. We compared 5 skills on formats covered, license, the agents they work in and upkeep; compared on 2026-10-01, and the table shows the date each fact was last read.

PPTX (Anthropic): PowerPoint decks inside Claude

PPTX is Anthropic's agent skill for PowerPoint files, for people who make or edit decks in Claude.ai, Claude Code or the Claude API. Its strength is template work: it unzips your .pptx or .potx, builds new slides from the template's layouts with bundled scripts, then renders every slide as an image and fixes overflow and overlaps before handing the file back. New decks are written with the pptxgenjs library, guided by a list of mistakes that corrupt files, such as writing a hex color with a leading # sign. Its limit is the license: the LICENSE.txt in its folder, read on 2026-10-01, forbids keeping copies outside Anthropic's services, making changed versions and redistributing it, so it cannot follow you to another agent. The folder was last changed on 2026-07-17, and the anthropics/skills repository had 179,254 GitHub stars on 2026-10-01. In Claude Code it installs with the document-skills plugin. Details: PPTX on CookiesForAI.

DOCX (Anthropic): Word documents and redlines inside Claude

DOCX is Anthropic's agent skill for Word documents and templates (.docx and .dotx), for Claude users who need reports, letters or contract markups as real Word files. Its strength is redlining: Claude edits the document XML directly, and bundled Python scripts check that every edit is recorded as a tracked change, accept tracked changes, add comments anchored to text and validate the file against the Office XML schemas. After writing, Claude renders the pages as images and looks at them. New documents are built with the docx library for JavaScript, with notes on traps such as A4 being its default page size. Its limit is twofold: the same proprietary license as PPTX, read on 2026-10-01, and a narrow trigger, since the skill says it is not for PDFs, spreadsheets or Google Docs. The folder was last changed on 2026-07-17. Details: DOCX on CookiesForAI.

XLSX (Anthropic): Excel files with working formulas inside Claude

XLSX is Anthropic's agent skill for spreadsheets (.xlsx, .xlsm, .xltx, .csv and .tsv), for Claude users who build financial models, clean exports or add columns to a workbook. Its strength is that formulas stay live: Claude writes formulas instead of pasting computed numbers, then a bundled script recalculates the workbook with LibreOffice and names every cell with an error, and the file is not delivered until that count is zero. When it edits your file, it follows your conventions and leaves existing formulas alone. Its limit comes from that same check: LibreOffice cannot evaluate newer Excel functions such as XLOOKUP and FILTER, so the skill avoids them and uses INDEX and MATCH. The license, read on 2026-10-01, again keeps it inside Anthropic's services. In our test on 2026-10-01, five of six AI assistants asked about agent skills for PDFs and spreadsheets pointed to Anthropic's document skills; ChatGPT described generic skill types instead. Details: XLSX on CookiesForAI.

PDF (Anthropic): every common PDF job inside Claude

PDF is Anthropic's agent skill for PDF files, for Claude users who need to pull data out of PDFs or change them. Its strength is breadth: one skill covers text and table extraction, merging, splitting, rotating, watermarks, new PDFs, encryption, image extraction and OCR of scanned pages, and filling in forms has its own guide with scripts that find fillable fields, fill them and render the pages to check the result. Its limit is that the main file reads as a cookbook of Python libraries and command-line tools: outside form filling, it does not require the agent to render and inspect each page of a new PDF, which OpenAI's skill does. The license, read on 2026-10-01, matches the other three: use only within Anthropic's services. Its folder was last changed on 2026-02-06, five months before the other three. Details: PDF on CookiesForAI.

PDF (OpenAI): the PDF skill you may use in any agent

OpenAI's pdf is an agent skill for creating and reviewing PDF files where layout matters, for anyone who works outside Claude or wants a skill they may copy and change. Its strength is a strict visual check: the agent builds PDFs with reportlab, renders pages to PNG images with Poppler after every meaningful change, and may not deliver until the latest render shows no clipped text, overlaps or broken tables. The skill folder carries its own Apache-2.0 LICENSE.txt, so use at work, changes and redistribution are allowed. Its limits are scope and upkeep: it does not cover form filling, merging, splitting or OCR, and OpenAI marked the openai/skills repository deprecated on 2026-06-22, pointing to openai/plugins instead. The skills CLI still listed it on 2026-10-01; the folder was last changed on 2026-02-02, and the repository had 27,835 stars that day. It needs Python and Poppler. Details: OpenAI's PDF skill on CookiesForAI.

Which of these you may use outside Claude

The license, not the feature list, decides most choices here. Anthropic's README calls its four document skills "source-available, not open source", and each folder's LICENSE.txt puts use under your agreement with Anthropic and forbids extracting the files from its services, keeping copies, making derivative works and redistributing them. You may read them on GitHub to learn how a complex skill is built; you may not install them into Codex, Cursor, Gemini CLI or any other agent. The open skills CLI does list them, and in our 2026-10-01 test two assistants suggested that route: Grok gave the npx skills add anthropics/skills --skill pdf command and named Codex and Cursor as compatible, and Google AI Mode said the document skills work in Cursor, GitHub Copilot and Codex. Both answers ignored the license.

Outside Claude, the only skill in this list you may use is OpenAI's pdf. For Word, PowerPoint and Excel in other agents, we found no openly licensed skill from a large publisher, or above our 5,000-star bar, on 2026-10-01. Small MIT-licensed skills exist, such as academic-pptx-skill (1,100 stars), but it describes itself as working alongside Anthropic's PPTX skill rather than replacing it. OpenAI's newer openai/plugins repository has a reports-pdfs-and-slide-automation skill for Codex, but neither its folder nor the repository had a license file on 2026-10-01, so its terms for use elsewhere are unstated. If your "documents" are Markdown notes rather than Office files, Obsidian Markdown is MIT-licensed and works in Claude Code, Codex and OpenCode.

How we compared

We read each skill's SKILL.md, its LICENSE.txt and its repository README on GitHub on 2026-10-01, and took stars, last-change dates and install routes from our item pages, checked the same day. We compared five things: which file formats each skill covers, what the license allows, which agents the publisher documents, whether the agent checks its own output, and whether the source is still maintained. We also used the answers to question 3 of our field study of six AI assistants, run on 2026-10-01. We did not run the skills on test files, did not measure output quality, and did not give legal advice: read the license yourself before you rely on any skill at work.

How to choose

• If you work in Claude.ai, Claude Code or the Claude API, install Anthropic's document-skills plugin: it brings all four formats, and on Claude.ai paid plans the skills are already available.
• If you work in Codex, Cursor, Gemini CLI or another agent and need PDFs, use OpenAI's pdf skill; it is Apache-2.0.
• If you need PDF form filling, merging or OCR outside Claude, none of the skills here covers it; ask your agent to use a library such as pypdf directly, or write your own skill.
• If you want to publish or sell a changed version of a document skill, start from OpenAI's pdf, never from Anthropic's four.
• If your notes live in an Obsidian vault rather than in Office files, use Obsidian Markdown instead of any of these.