Claude for Word, Excel and PDF: Real Files, Not Markdown

Ask Claude for a report and, out of the box, you get markdown in a chat window. It looks fine on screen. Then you paste it into Word and spend the next forty minutes rebuilding headings, fixing the table that came through as pipes and dashes, and creating a table of contents by hand. Most people conclude that this is simply what AI does: it writes text, and formatting is your job.

That conclusion is wrong, and the fix costs nothing. With the right skills installed, Claude produces actual files. A .docx that opens in Word with real heading styles and a clickable table of contents. An .xlsx where the totals are live formulas, not pasted numbers. A .pptx you can present from. A filled PDF form. You ask in plain English, and a finished file lands in your folder.

We run SkillProof, a catalog where every skill gets installed on a clean machine and tested against real office work before we publish a verdict. This guide covers the document skills specifically: what each format can do, what our tests actually produced, where things break, and how to set the whole stack up in about five minutes.

The part most people miss

Here is the detail that surprises nearly everyone outside the developer crowd: the document skills are not a community hack. They are Anthropic’s own flagship skill set, published in the official anthropics/skills repository, and they are the most polished skills we have ever tested.

Our documents category tells the story in numbers. The four core skills, DOCX, XLSX, PPTX and the PDF toolkit, all passed testing with perfect install and trigger scores. Three of the four scored 9/10 on output quality; PowerPoint scored 8. For context, roughly half the community skills we pull from GitHub fail on first install. These four worked immediately, every time, on a clean machine.

The mechanism is worth one paragraph, because it explains both the power and the limits. A skill is a folder of instructions that teaches Claude a procedure (the full picture is in what are Claude skills). The document skills teach Claude to write and run small scripts that build Office files properly: real OOXML under the hood, real PDF objects, not an export of chat text. That is why the output opens cleanly in Word, Excel, PowerPoint, Google Docs and LibreOffice. Claude is using the same file formats those applications use, from the inside.

So the question shifts from “can Claude make a Word document” to “how good is it per format.” That is where our test notes come in.

Word: the tracked changes surprise

The DOCX skill is the one to install first, and our test run shows why. We asked for a long-form business report from raw notes. The result was a 14-page document with a table of contents, numbered headings, styled tables and a letterhead. Not markdown dressed up. The TOC was a live Word field that updates when you add sections. The headings used Word’s actual style system, so changing “Heading 2” once restyled the whole document.

The feature that changes daily office life, though, is tracked changes. We handed Claude an existing contract and asked for edits. It returned the file with proper Word revisions, the red strikethrough and margin balloons your legal counterpart expects, without corrupting anything else in the document. Comments work too, both reading and writing them.

Think about what that covers in a normal week. Status reports that used to be a formatting chore. Memos on letterhead. A first-pass redline of a vendor agreement before the lawyer sees it. Reorganizing a 30-page document some committee wrote by accretion. In each case you describe the outcome and receive a file, and the file behaves like a colleague made it in Word, because structurally it was made the same way.

Two practical notes from testing. Be explicit about structure up front (“two-page memo, letterhead, three sections”) because Claude follows a spec better than it guesses one. And keep source material in the project folder so Claude can read it directly instead of you pasting fragments.

Excel: formulas, not frozen numbers

The failure mode of AI spreadsheets is subtle. Plenty of tools will generate a table where cell C10 contains the number 4,820. It looks like a total. It is actually a corpse: change any input and it sits there, wrong forever. The XLSX skill writes =SUM(C2:C9) instead, and that one difference decides whether the file is a deliverable or a screenshot.

Our test was deliberately ugly. We gave Claude a 40,000-row export with malformed headers, the kind of file a billing system produces when nobody has looked at the template since 2019. The skill cleaned the structure, added computed formula columns and built a pivot-style summary sheet on top. Then the real check: we opened the file in desktop Excel and in Google Sheets and changed input values. The formulas recalculated correctly in both.

That recalculation test matters more than any demo. It means the spreadsheet Claude hands you is a living model. Your finance person can audit the logic by clicking into cells, exactly as they would with a junior’s work. It also means Claude can inherit your existing files: it reads .xlsx, .xlsm, .csv and .tsv, so “clean up the export in my downloads folder and add a margin column” is a one-sentence request.

One habit worth forming: tell Claude which cells should be formulas and which should be hardcoded assumptions. It defaults sensibly, but a model where the assumptions block is clearly separated from the calculation block is easier to defend in a meeting.

PowerPoint: from brief to deck

The PPTX skill turns a brief plus data into an actual presentation file. In our test we gave it a CSV and a written brief and asked for a quarterly business review deck. It produced 12 slides with charts generated from the data, consistent styling and, usefully, speaker notes under each slide. The chart numbers were accurate against the source CSV, which we verified cell by cell.

It was also the only skill of the four that needed a second pass: one table came out with cramped column widths and we asked for a fix, which took a single follow-up message. That is why it carries an 8 rather than a 9 in our scoring, and it is an honest picture of the experience. Expect a strong first draft and one round of visual touch-ups, not pixel-perfect output on the first try.

The right mental model is “deck builder,” not “designer.” Structure, charts, notes and consistency are what it does well. If your company has a heavily branded template with custom masters, you will still spend time in PowerPoint afterward, though far less than starting from a blank slide.

PDF: the format everything ends up in

The PDF toolkit is less glamorous and possibly the most broadly useful of the set, because PDFs are where office documents go to become someone else’s problem. Our test session covered three jobs. Claude merged six separate PDFs into one file. It filled in a government form, the genuinely fiddly kind with named fields. And it extracted tables from a scanned invoice using OCR.

That last one deserves a caveat straight from our test notes: OCR quality depends on scan quality. A clean 300dpi scan came through accurately; expect a crumpled phone photo of a receipt to need human review. But for the everyday cases, pulling numbers out of supplier invoices, splitting a 200-page PDF into chapters, watermarking a proposal, encrypting a file before sending it, the skill quietly replaces a whole shelf of paid web tools, along with the awkward moment of uploading a confidential contract to smallpdf-whatever.com. Everything runs on your machine.

If your work involves receiving documents rather than producing them, install this one first and DOCX second.

Three workflows we actually run

Individual file tricks are nice. The compounding value shows up when you chain them. These three scenarios come from our own testing and from readers who wrote in.

The monthly report pipeline. Inputs: a CSV export from your analytics or billing system, plus last month’s report as a reference. The request: “Read this export, update the metrics workbook with formulas, then write the monthly report in the same structure as the attached document, as a .docx with our letterhead.” Claude cleans the data in Excel, computes the month-over-month changes as live formulas, then writes the Word report referencing the real numbers. What used to be a first-Friday-of-the-month half day becomes a twenty-minute review of a draft. The review matters, and we will get to that, but the assembly work is gone.

Contract redlining. A vendor sends their standard agreement. You keep a short file of your negotiating positions: liability caps, payment terms, termination notice. The request: “Review this contract against our positions and propose edits as tracked changes.” Claude returns the .docx with revisions and margin comments explaining each change. Your lawyer, or you, then accepts or rejects in Word as normal. To be clear about the boundary: this is a first pass that catches the standard asymmetries, not legal advice. It makes the expensive hour with counsel about judgment instead of typing.

Deck from bullet points. You have meeting notes or a rough outline and a deck due at 3pm. Request: “Turn these bullets into a 10-slide deck for an executive audience, one idea per slide, with speaker notes I can rehearse from.” The pptx skill handles layout and pacing; you spend your remaining time on the argument rather than on aligning text boxes. If there is data involved, attach the spreadsheet and ask for charts instead of tables. Executives read charts.

Where it breaks, honestly

We publish failures as prominently as passes, so here is the list of things the documents stack does not do well.

Complex layouts. Multi-column magazine-style layouts, text wrapping around images, intricate brand templates with custom masters. The skills handle standard business formatting well and creative layout poorly. If your document’s value is visual design, a human still owns that.

Huge files. The 40k-row spreadsheet passed, but performance degrades as files grow, and a 500MB workbook with fifteen years of cruft will time out or crawl. Same for a 1,000-page PDF. The practical ceiling is generous for office work and real for data-warehouse work. Split big jobs.

Fonts. Claude builds files with the fonts available in its environment. If your corporate template requires a licensed font like a custom brand face, the file will open with a substitute until someone with the font touches it. Standard fonts (Calibri, Arial, Times) are safe; boutique typography is not.

Scanned inputs. As noted, OCR is only as good as the scan. Numbers extracted from poor scans need a human check before they enter anything that matters.

The blank-check request. “Make me a report” with no source material produces confident, generic filler in a beautifully formatted file. Nice formatting makes weak content easier to miss, which is a new failure mode to watch for. Give it real inputs.

None of these are dealbreakers for the target use case, which is the ordinary document work of an ordinary office. They are the edges of the envelope.

The junior analyst question

Someone always asks whether this replaces the junior person who used to build these files. Our stance, having run the tests: it replaces the typing, and it does not replace the person, but it does change what the person is for.

The stack genuinely absorbs the mechanical layer, formatting reports, updating workbooks, assembling decks, first-pass redlines. If a role consists entirely of that layer, the role is exposed, and pretending otherwise helps no one. What it does not absorb is accountability. Claude will build a flawless model on a wrong assumption without blinking. Somebody still has to know that the Q2 number looks off, that this vendor always hides the auto-renewal in section 11, that the CFO wants the risks slide before the asks slide. That judgment used to be learned by doing the typing. Teams now have to teach it deliberately, by making juniors review and challenge generated drafts rather than produce them from scratch. The reviewers who catch the wrong assumption become more valuable, and the throughput expectation for everyone rises. That is the honest trade.

Installing the documents stack

The short version, since the full install guide covers every surface and failure mode:

git clone https://github.com/anthropics/skills
cp -r skills/document-skills/docx ~/.claude/skills/
cp -r skills/document-skills/xlsx ~/.claude/skills/
cp -r skills/document-skills/pptx ~/.claude/skills/
cp -r skills/document-skills/pdf ~/.claude/skills/

Restart Claude Code and test the trigger without naming any skill: “turn these notes into a two-page Word memo.” If a .docx appears, you are done. On claude.ai and the desktop app, the same skills upload as zips under Settings, no terminal required. Our current rankings for the whole category, including the community skills that build on this stack, live at best document skills.

FAQ

Can Claude really create a Word document I can open and edit?

Yes. With the DOCX skill installed, Claude writes genuine .docx files with real Word styles, tables of contents, page numbers and tracked changes. In our test it produced a 14-page report with a working TOC that opened cleanly in Word, Google Docs and LibreOffice. You edit it like any other document, because structurally it is one.

Does Claude put working formulas in Excel files or just numbers?

With the XLSX skill, working formulas. We verified this directly: formula columns and summary calculations recalculated correctly when we changed inputs in both desktop Excel and Google Sheets. You can also ask which values should stay as fixed assumptions, which is good modeling practice anyway.

Do I need to be a developer to use these skills?

No. Installing means copying four folders (or uploading zips on claude.ai), and after that everything is plain English: “fill in this PDF form,” “clean up this export,” “redline this contract.” The one-time setup is the only step that touches anything technical, and the guide walks through it click by click.

Can Claude fill in PDF forms and read scanned documents?

Both, with the PDF toolkit. It filled a government form with named fields correctly in our tests, and it extracts text and tables from scans via OCR. The caveat is scan quality: clean scans extract reliably, bad ones need human review. It also merges, splits, watermarks and encrypts PDFs locally, with no upload to third-party web tools.

Which document skill should I install first?

If you mostly produce documents, DOCX; it has the broadest daily payoff and the tracked-changes feature alone justifies it. If you mostly receive documents, the PDF toolkit. Since all four install in the same five minutes and cost nothing beyond your existing Claude plan, the realistic answer is the whole stack at once.