Getting work out of Whizi as real files

The short answer

Ask a model in Whizi for a spreadsheet or a document and you get a downloadable file card rather than a formatted table stuck inside a chat message. The file is generated in your browser as a genuine .xlsx or .docx that opens in Excel, Numbers, Word or Google Docs.

This is the difference that matters against copying a markdown table out of a chat window: the columns are real columns, numbers are real numbers rather than text, and a multi-sheet workbook arrives as a multi-sheet workbook.

Asking for a spreadsheet

Be explicit that you want a file. "Put this in a table" gets you a table in the message; "give me this as a downloadable spreadsheet" gets you a file card.

Three things worth specifying in the same request, because they are far easier to state up front than to fix afterwards:

• The columns you want, by name. Otherwise you get the model idea of a sensible schema, which is usually close and never exactly right.
• What to do with missing values. Say "leave the cell empty rather than guessing" if the source is incomplete, because inventing plausible values is the failure mode here.
• Separate tabs, if you want them. A single request can produce a multi-tab workbook, so "one tab per region" or "raw data on one tab and the summary on another" both work.

The strongest input for this is an uploaded .xlsx or .csv rather than a PDF, because the model receives real cell structure rather than extracted prose. Spreadsheet in, spreadsheet out is by some distance the most reliable version of this workflow.

Documents, charts and HTML

Documents. The same approach produces a real .docx with headings, paragraphs and tables. Useful when the output is going to someone who expects a document rather than a chat log: a summary of a report, a first draft of a policy, meeting notes structured into sections.

Charts. When a model produces a chart, it renders as an actual chart in the conversation, in bar, horizontal bar, line, area or pie form, and you can download it as an image. That makes it usable in a deck or a document without rebuilding it.

HTML. Generated HTML renders in a preview pane alongside the code, so you can look at the thing before deciding whether to use it, and switch between the rendered view and the source. It can also be downloaded.

Check the numbers

A generated spreadsheet looks authoritative in a way a chat message does not, and that is exactly the risk. The formatting being correct tells you nothing about whether the figures are.

Two checks that catch most problems in under a minute. First, spot-check three or four values against the source, chosen from different parts of the data rather than the top. Errors cluster where the source was ambiguous, which is rarely row one. Second, check the row count. A silently truncated extraction is the most common failure, and it is invisible unless you count.

If the data matters, the cheap verification is to ask a second model to extract the same thing from the same source and compare the two files. Where they agree, confidence is reasonable; where they diverge is exactly where you should look. See comparing models side by side.