The blank box problem

The situation

It is Thursday, 07:40. You open a new chat to draft the weekly note to branch managers. Before you can ask for anything you type the same eleven lines you typed on Monday, and the Monday before that: you cover fourteen branches, the readers are branch managers and not head office, "unit" means a pallet, keep it under 250 words, do not state a number that is not in the export.

Then you ask for the note. It comes back usable. Nine minutes, six of them spent before the request.

On Friday you are doing it from your phone, so you skip four of the eleven lines. The draft comes back at 600 words, pitched at head office, with a figure in it that appears nowhere in your data. You do not notice the figure until a branch manager emails to ask where it came from.

Nothing malfunctioned. The box was blank, you filled it differently, and you got a different answer.

The principle

The context you retype from memory under time pressure is the context that gets dropped, and what gets dropped is what the tool fills in for you.

Nothing you depend on carries between sessions unless you deliberately put it somewhere that does — and whatever a tool carries on its own is the part you did not choose. So the most durable part of your work, the part that has not changed in two years, gets retyped by you, from memory, in the ninety seconds before a meeting. Friday's invented figure is what one of those omissions looks like from the outside.

Four buckets cover almost everything worth saying twice. Who you are — your role, your readers, and the decision the output feeds. What good looks like — length, structure, tone, and the test for finished. What the recurring inputs are — the same export, the same report, and the house words that mean something specific in your organisation. What must never happen — the prohibitions.

Three of the four are easy to write. The fourth is the one nearly everyone skips, and it is the only one that constrains the tool when your request is vague.

Worked example

Shown here in Gemini.

Two openers from the same week, by the same person, for the same task:

Monday 08:12
I run ops for 14 branches at a building-products distributor. Write the
weekly note to branch managers from the attached export. Under 250 words,
plain text, movements first. "Unit" means a pallet. Don't use any number
that isn't in the export...

Friday 07:44
Write the weekly note, keep it short, numbers attached.

The overlap between those two is the context. Extracted, it looks like this:

WHO I AM
  Regional ops lead, 14 branches, building-products distribution.
  Readers are branch managers. They act on this the same day.

WHAT GOOD LOOKS LIKE
  Plain text, under 250 words. No preamble, no closing offer to help.
  Movements first, then one line headed "Needs a decision", or nothing.

RECURRING INPUTS
  The Monday export: one row per branch, columns booked / picked / short.
  House words: branch = location, account = customer, unit = pallet.

NEVER
  Never state a number that is not in the export.
  Never name an account in anything a branch manager can forward.
  Never soften a shortfall into "slightly behind" — give the figure.

Nothing in that block will be false next Thursday. That is the whole reason it is worth writing down once. Where it goes, and which of your lines do not belong in it, is the next lesson.

In your tool

Before deciding where to put the block, know what your tool already carries between chats on its own — because whatever it carries by itself is the part you did not choose.

Product surfaces checked 2026-08-04.

Try it

Open your three most recent AI chats. Copy the first message of each into one blank document. Mark every sentence that appears, in some form, in more than one of them. Sort the marked lines into the four buckets above.

You are done when every bucket has at least one line and the NEVER bucket has at least two. Do not polish the wording. An empty bucket is the finding: it is the gap that explains the outputs you had to rewrite by hand.

Common failures

• The growing preamble. Every disappointing output adds a sentence to the top of the next request, and nothing is ever deleted. Four weeks later the preamble runs to 400 words, the prohibition that matters sits between two tone requests, and it gets weighted like everything else around it.
• Context made only of nouns. "We are a 200-person distributor with fourteen branches" is background, not a standard. It says what you are and nothing about what a finished output looks like, so what comes back is well-informed and still the wrong shape.
• No prohibitions. People record what they want and never what is forbidden. The prohibition list is the only part of your context that does work when the request is ambiguous — which is exactly the moment you were relying on it.