Work

A browser assistant can describe your spreadsheet. Work opens it, fixes it, and saves it — then hands you the report, the deck, or the PDF.

Real files, no Office required

Work reads and writes genuine Open XML — .docx, .xlsx, .pptx — by handling the format directly. Microsoft Office does not need to be installed, on any platform.

Research that remembers

The failure mode of research is not finding too little. It is finding something, mentioning it in passing, and losing it — so the same page is opened again twenty minutes later to recover a number that was already known.

Work writes findings down as it goes. Each records the claim, the page it came from, the wording that supports it, and the date it was checked.

Three things follow from writing it down:

• Nothing is fetched twice for the same fact. From the second read of a page onwards, Work is handed what it already noted from it before the page itself.
• Findings survive a long run. A conversation gets compacted as it grows; the notebook is on disk and does not.
• The same claim from a second source is a second finding. Re-reading one page does not turn one fact into two, but corroboration from elsewhere is recorded as exactly that.

References that resolve

Write the prose and cite findings inline as [f1]. Work resolves each to a numbered reference, ordered by where a reader first meets it, and builds the source list to match:

Devin bills by ACU [1]. Cursor is $20 [2], or $40 a seat [2].

## Sources
1. https://devin.ai/
2. https://cursor.com/pricing

Two findings from one page share one number, because they are one reference. The list includes only pages something was actually cited from — a page that was opened and yielded nothing is not a reference, and padding a bibliography with it would misrepresent the work.

A citation matching no finding is reported and left visible in the text as [f9?] rather than quietly dropped. A report whose references do not resolve is worse than one that admits it, and only the writer can fix it. A cited finding with no source URL renders as [unsourced] and gets no number: true things somebody said are not things a reader can check.

In the desktop app the notebook sits beside the transcript, grouped by topic, with each source as a link — so a claim can be checked against its page before it becomes a paragraph in something you sign. Pages read more than twice are listed separately, because that is the shape of a run that was not writing things down.

Decks

Research usually ends in something you present. Ask for a .pptx and Work builds one from Markdown:

# The problem

- Stock leaves without anyone noticing
  - Worst at month end

> Say this part slowly

One slide per heading. Bullets keep their nesting from the indentation. A > quoted line becomes that slide's speaker notes rather than something the audience reads.

Checking its own work

Work reopens what it produced and checks the expected content is really there before telling you it is done. That catches a missing section, a dropped figure, or a file that never saved.

Email and Drive

Gmail and Google Drive, on every platform. Connect once with /connectors connect google and Work can read your Drive and your mail, save what it makes into Drive, and draft replies you review before they go — “read the Q3 brief in my Drive, turn it into a deck, and put it back” is one request. See Connectors.

Outlook, on Windows. Inbox triage, reading a thread and drafting a reply through your own Outlook profile. Drafts are saved, never sent; sending is a separate action that always asks for confirmation, regardless of your other permission settings.

From a conversation to a document

Talk something through in /chat, then /work write that up as a report — Work starts with the conversation's conclusions already in hand.

A scoped workspace

Work operates inside a folder you choose, which keeps it away from your code projects. Providers, budgets, policy and the audit log are shared with Build and Motion — one set of controls, three faces.