The one-sentence rule

The situation

Thursday, four o'clock, two things left on the list. The first: cut the service-description section of a statement of work from 900 words to 400 without breaking the clause numbering. The second: find out which of the fourteen signed agreements in the shared drive renew automatically, and when notice is due on each. You do both in whichever tab is already open. The first comes back as clean prose you then spend fifteen minutes pasting in and renumbering. The second comes back confident and wrong, because the assistant could only see the one contract you had in front of you. Neither tool malfunctioned. You routed both tasks by habit.

The principle

Work that lives inside one document belongs to the assistant inside that document, and work that spans files, tools or steps belongs to the general assistant.

This is not a claim about which model is smarter. It is a claim about two things no model can guess for you: how many sources the answer needs, and where the finished work has to land. An assistant built into a document can see one file and change it in place — that is its whole advantage and its whole limit. A general assistant can hold fourteen files, a standing instruction and a multi-step task at once, and it hands back text that still has to be moved somewhere.

Say the rule out loud before you open anything: one document, or across the work. Two questions settle almost every case in about three seconds. How many sources does the answer need? Where does the finished thing have to end up?

Worked example

Shown here in Claude.

Route both Thursday tasks on paper first.

Task A — Cut the service description in SOW-2026-114 to 400 words.
  Sources the answer needs: 1
  Where the finished work lands: back in that file, numbering intact
  Route: inside the document

Task B — Which of the 14 signed agreements auto-renew, and when is notice due?
  Sources the answer needs: 14
  Where the finished work lands: a new table that lives in none of them
  Route: across the work

Task B, assigned to a Project holding all fourteen agreements:

Job: Build the renewal-notice table from the agreements in this project.

Inputs: All 14 signed agreements. Ignore drafts and redlines.

Output: One row per agreement — counterparty, renewal date, notice window
in days, and the clause number you read it from.

Standard: If an agreement does not state a notice window, write "not stated"
and give the clause you checked. Do not infer one.

The clause-number column is doing the work of a citation. It is what makes a fourteen-file answer checkable back in the single file it came from. Task A never needed that column, because you were looking at the file the entire time — which is the same fact the rule is built on, stated from the other side.

In your tool

The rule is stable. The surfaces move. In any ecosystem you are looking for two things: an assistant that can edit the file you are already in, and an assistant that can see across files.

Product surfaces checked 2026-08-04.

Two catches worth naming. In the left column, a side-by-side panel and an in-file assistant are not the same thing: the first hands you finished text, the second changes the document. In the right column, an assistant that reaches across your files reaches only across its own ecosystem's files — a four-source task whose fourth source is a PDF in someone's inbox is a copy-paste job in every tool on this list.

Try it

Open your calendar and list the last five things you used an AI tool for. Beside each, write two answers and one word: how many sources the answer needed, where the result had to land, and the route — document or work. Five minutes, nothing open.

You are done when you can point at one task you routed the wrong way last week and say the sentence that would have caught it.

Common failures

• The open-tab rule. The task goes wherever you already are. Routing is a decision, and a decision made at four o'clock on a Thursday defaults to whatever is on screen.
• Asking the in-document assistant a question the document cannot answer. It can see one file, so it answers from one file — confidently, in exactly the tone it uses when it is right. This looks like the model inventing things. It is a routing error wearing a costume.
• The paste loop. File out to the general assistant, edited text back in, comments and formatting lost, then a second round to repair what the paste broke. It happens because the first answer was good. The cost is all in the return trip, and nobody counts the return trip when choosing.