Document generation from scattered sources

If file operations are the demo that sells Cowork, document generation from scattered sources is the use case that makes people keep their subscription. For most knowledge workers — PMs, consultants, marketers, finance leads — a significant portion of every week is spent assembling information from multiple places into a single output. Cowork doesn't make that faster. It does it for you.

The canonical example is Anthropic's own demo: expense report from receipt photos. You drop a folder of photos into raw/, run a single prompt, and get a formatted expense report in out/. Not a list of extracted figures you paste into a template. An actual spreadsheet — with totals, categories, and the structure your finance team expects — produced from images Cowork read and understood.

That distinction matters. The output isn't text you copy-paste into Office. It's an actual .xlsx with working formulas, or an actual .docx with proper headings and tables, or an actual .pptx with speaker notes and slide layouts. Cowork produces finished files, not raw material for finished files.

The use cases that generate the most ROI

Expense reports from receipt photos. The canonical demo. Works on photos, scanned PDFs, and digital receipts mixed together in a single folder.

Status reports from project files. Point Cowork at a folder of meeting notes, Slack exports, and planning docs. Ask for a weekly status report. It reads across the sources, identifies progress and blockers, and produces a structured document — in your organisation's format if you give it a template.

Weekly analytics dashboards pulling from multiple sources. Export CSVs from your analytics tools, drop them in raw/, and ask for a summary dashboard in out/. The result is a multi-tab Excel workbook with actual VLOOKUPs and conditional formatting, not a markdown table.

Competitive landing-page audits. Cowork paired with Claude in Chrome navigates competitor sites, takes screenshots, and reads the page content. The output: a side-by-side Word document comparing your positioning to each competitor's, with a gap analysis section. This is research that would take a half-day manually; Cowork does it while you work on something else.

Newsletter DNA extraction. Drop your past twelve newsletter issues into a folder. Ask Cowork to produce a voice and structure profile — what topics you cover, how you open, how you close, what sentence-level patterns define your writing. The output becomes your permanent voice file, which you reference in every future content task.

The prompting pattern

Structure document-generation prompts the same way every time:

"Read everything in raw/. Produce [specific output file] in out/. Include [sections]. Log any data you couldn't extract in out/open_questions.md."

The key elements:

• Name the input folder (raw/) so Cowork knows where to read from
• Name the specific output file (out/expense-report.xlsx) so it knows what to produce and in what format — the file extension drives the skill it uses
• Specify the sections so the structure matches what you need, not what Cowork guesses
• Add the open-questions log so data gaps surface explicitly instead of being silently filled with assumptions

The open-questions log is especially important for financial documents. If a receipt is illegible, you want out/open_questions.md to say "Receipt dated March 4 — merchant unreadable," not a spreadsheet row with a guess.

Why the output format matters

When you ask for out/report.xlsx, Cowork loads its xlsx skill and produces a real Excel file — multi-tab structure, working VLOOKUP formulas, conditional formatting where appropriate. When you ask for out/brief.docx, it loads the docx skill and produces a real Word document with a table of contents, headings styled to Word's heading hierarchy, and page numbers.

The file extension in your output specification isn't cosmetic. It tells Cowork which skill to use, and those skills govern every structural and formatting decision in the output. If you ask for out/report.md, you get markdown. If you want a spreadsheet, say so explicitly.

The multi-source advantage

The reason Cowork beats Chat for document generation isn't intelligence — it's breadth. Chat's context window fills up. If you upload five PDFs, you're at 30% context before you've typed a word. At ten documents, you're managing what gets included and what gets truncated. At twenty, Chat breaks down.

Cowork reads the full folder. It processes documents sequentially, references earlier findings later in the task, and synthesises across the entire set without you managing the context. That's what makes research-heavy document generation viable rather than frustrating.