TPEE Toptronic® Prompt Engineering Editor

114. Revisiting Context 1 - Context Is a Budget, Not a Container

Stop treating the window as storage: every token competes with every other token for the model's attention.

By Jacques Botte, founder of Toptronic®. Last updated 17 October 2026.

The lesson

The first pass at context engineering taught you to fill the window with the right material. The second pass starts from a harder fact: the window is a BUDGET, and the model does not read it the way a person reads a folder. Attention is finite and it is shared. A long document in the middle does not merely sit there; it competes with your instruction at the top and your format request at the bottom, and the competition is not fair.

Two consequences follow immediately. First, irrelevant material is not neutral - it is a cost paid in attention, and models measurably lose track of instructions buried in long inputs. Second, more context is not more information. A thousand tokens of restated background add nothing the model did not already have from the brief, and they push the actual question further from the point where attention is strongest.

The practical reframe is to budget tokens the way you budget money: a total, a split, and a rule for what gets cut first. A working split for a document task is instruction and format, then the material that must be quoted exactly, then a compact summary of everything else, then nothing. When the material exceeds the budget, you compress or select - you never simply overflow and hope.

Learn the shape of the window as well. The beginning and the end are read most reliably; the middle is where an instruction goes to be forgotten. So state the task at the start, restate the acceptance check at the end, and keep the long material in the middle where losing detail costs least. This is not a trick, it is the same reason a good report has an executive summary and a conclusion.

And remember what the window is not: it is not memory and it is not a database. It is the working set for one turn. Anything that must survive the turn belongs somewhere on disk, in a file the next turn explicitly reads, so that what enters the window is a decision rather than an accumulation.

Check yourself

Question 1: Why is irrelevant material in the window not neutral?
  1. It costs money only
  2. It competes for finite attention and can push the real instruction out of the strongest reading position — correct
  3. It makes the model slower but not worse
  4. It is ignored completely

Answer: It competes for finite attention and can push the real instruction out of the strongest reading position

Attention is finite and shared. Filler is a cost paid by the instruction you actually needed followed.

Question 2: Where in a long input is an instruction most likely to be lost?
  1. The very beginning
  2. The middle — correct
  3. The very end
  4. It is equally reliable everywhere

Answer: The middle

Beginning and end are read most reliably. State the task first, restate the acceptance check last, keep bulk material in the middle.

Question 3: What is the window, honestly described?
  1. Permanent memory
  2. A database
  3. The working set for one turn — correct
  4. A backup of the project

Answer: The working set for one turn

Anything that must survive the turn belongs in a file the next turn explicitly reads, so what enters the window is chosen.

← Previous lesson · All 134 lessons · Next lesson →

The full course — 134 lessons and 402 quiz questions — ships inside the app. Get TPEE to study it offline.

This lesson is the written summary of a course that ships inside the TPEE prompt editor for Windows - thirteen sections, a 0-100 score before you send anything, and it works with no network connection. The three one-time prices, and a 14-day free trial.