Illustration showing unstructured source material flowing through an AI workspace into a spreadsheet, report and presentation.
Illustration showing unstructured source material flowing through an AI workspace into a spreadsheet, report and presentation.

Move beyond chat and into working files

It is easy to think of AI as something you ask to draft text. The more useful shift is letting the model build the actual files where the work will continue.

Claude can now create and edit Excel workbooks, PowerPoint decks, Word documents and PDFs inside a conversation. It does this with code running in an isolated environment, so the output can contain working formulas, multiple sheets, charts and editable slides — rather than a table or outline you still have to rebuild yourself.

Anthropic’s current documentation confirms that file creation is available on the web, desktop and mobile across Free, Pro, Max, Team and Enterprise plans. Individual uploads and downloads are limited to 30 MB. The work counts against the normal usage limits of your plan, and generating files can consume more capacity than an ordinary chat.

The bigger opportunity: one source, several deliverables

The best use case is not simply “make a PowerPoint.” It is asking one analysis to produce a coordinated set of working materials.

Imagine you have a CSV export of customer feedback, meeting notes and a document describing the organisation’s goals. You could ask Claude to:

1. clean and organise the data in Excel,
2. add formulas and charts you can inspect,
3. write a report with findings and caveats,
4. build a short presentation using the same numbers and narrative.

Anthropic demonstrates a similar workflow in which survey data becomes a workbook, report and slide deck. My conclusion is that this is far more valuable than commissioning three unrelated AI outputs. Keeping the work in one conversation makes it easier to demand consistent figures, definitions and conclusions across every file.

Three practical examples

**A club or household budget.** Upload transactions and last year’s budget. Ask for a workbook with categories, monthly actuals, variance and a chart, followed by a concise report based on the workbook.

**Project reporting at work.** Provide an action list, risk log and meeting notes. Request an editable status report and five slides for a steering group, with every conclusion traceable to source material.

**A survey or course evaluation.** Upload anonymised responses. Ask Claude to group free-text themes, calculate distributions for rating questions and isolate responses that need human interpretation. Then have it create a report and presentation.

How to commission the work well

A strong prompt defines the deliverables, how they relate to each other and what must be verified. Specify the columns, metrics, charts and audiences you need. Ask for a raw-data sheet that remains untouched, plus a separate sheet documenting assumptions and formulas. State clearly what the model must not infer.

Ask Claude to list the files it created and run a final validation: do totals, date formats, percentages and headings agree? Then open the files in Excel, PowerPoint or their equivalents and test them. Change a few inputs to confirm that formulas and charts really update.

Use with care

A polished workbook is not necessarily a correct workbook. Review formulas, samples, units, tax treatment, currencies and rounding. Any output used for financial, legal or safety-critical decisions needs qualified human review.

Do not upload personal data, trade secrets or other sensitive material unless your organisation permits it. Team and Enterprise administrators can control the capability and restrict network access. Anthropic explicitly warns that external files or websites may contain hidden instructions intended to make the model disclose information. A cautious organisation should therefore begin with network access disabled and allow only approved domains when a task genuinely needs them. MCP connections require a separate security assessment.

Try it yourself

Start with non-sensitive material and reuse this prompt:

“Analyse the attached files and create three coordinated deliverables:

1. An editable Excel workbook with an untouched raw-data sheet, a cleaned-data sheet, an analysis sheet with clearly documented formulas, and an overview containing no more than three useful charts.
2. A short Word report covering method, key findings, uncertainties and recommended next steps.
3. A PowerPoint deck of no more than six slides for [audience].

Once the first version is ready, avoid vague feedback such as “make it nicer.” Give precise instructions: “make the chart scales comparable,” “move assumptions to a dedicated sheet,” or “reduce the deck to one decision per slide.” The second and third passes are often where AI-generated files stop being an impressive demonstration and become useful working material.