Main idea

A building-management context file passes a thousand lines, and you ask the agent to split it into as many files as there are topics. It does so, cleanly, and the result doesn't convince: the works and the calls for tender end up together, the debtors stay drowned in the record of people involved, a boundary falls where you saw none.

Nothing failed. Splitting means deciding what goes with what, and the instruction carried none of those criteria — neither "people on one side, debts on the other", nor "whatever gets consulted during a general meeting has to fit in a single file". Lacking them, the agent applies its own grouping logic, which is a logic of topical similarity. What it returns isn't a bad split, it is somebody else's split: the split of someone who has never had to open these files on a Monday morning to answer an email.

And the missing criterion wasn't available at the moment of asking. Faced with a file of several thousand lines, you no longer see the structure you'd want yourself — you recognize it by reading it, or you recognize that it isn't there. The first split therefore counts as a proposal to correct: it makes one possible organization visible, and it is by refusing it that you finally phrase the splitting directives you couldn't have written in cold blood. That is how a single file becomes fourteen thematic files in one session, then fifteen the same day, when debtor management turns out to deserve a space of its own.

Why it matters

This keeps you from reading the dissatisfaction as a failure of the model. The reflex in front of a disappointing split is to start again with the same instruction or to take the work back entirely; the useful gesture is to write down the grouping criterion, once, and let it govern the splits that follow.

It also says what you really delegate when you delegate a piece of shaping: not execution work, but a series of organizational decisions nobody will ask you for and that will then structure every use of the context.

Nuances and limits

Where nobody has a pre-existing use for the material, the model's structure is as good as the one you'd have produced, and arguing about the split only costs time. The rule only bites where someone will have to live inside those files.

And it has a flip side: splitting directives written too early freeze an organization the material hasn't yet justified, and produce empty files alongside files that overflow.

Open questions

• What, in a proposed split, signals that it is bad other than the discomfort of the person reading it?