TPEE Toptronic® Prompt Engineering Editor

115. Revisiting Context 2 - What Deserves to Be in the Window

A selection discipline: the five kinds of material worth spending tokens on, and the four that are not.

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

The lesson

The second pass at selection is a list, because a list can be applied under pressure. Five kinds of material earn their tokens: the INSTRUCTION (what to do, in the imperative), the ACCEPTANCE CHECK (how anyone will know it is done), the IRREDUCIBLE FACTS (the figures, names, identifiers and constraints that cannot be derived), the EXACT QUOTES (the passage to be edited, the error message, the schema), and ONE WORKED EXAMPLE of the output shape you want.

Four kinds of material almost never earn them. Restated background the model already has from the brief. Narrative of how the work went, which is a status report for a person, not an input for a task. Hedged commentary about what might be difficult. And anything included because it was close to hand - the whole file when one function was needed, the whole log when one error mattered.

The most common failure is the second-hand summary. Somebody sums up a document, the summary loses the exception that mattered, and the model answers confidently about a world that does not exist. If a detail is load-bearing, quote it exactly; if it is not, cut it. Half-remembered detail is worse than absent detail, because it looks like evidence.

Two mechanical tests settle most arguments. The DELETION test: remove the paragraph and ask whether the task could still be done correctly - if yes, it was decoration. The CITATION test: for each fact, name where it came from - if nobody can, it is a guess wearing a fact's clothes. Run both before sending, not after the answer disappoints you.

Finally, sequence the material for the reading order rather than the writing order. The order in which you gathered it is almost never the order in which it should be read. Put what must not be missed where attention is strongest, and let the bulk sit in the middle where a lost detail costs the least.

Check yourself

Question 1: Which of these earns its tokens in the window?
  1. A narrative of how the work went so far
  2. The exact error message, quoted — correct
  3. Hedged commentary about difficulty
  4. The whole repository

Answer: The exact error message, quoted

Exact quotes are one of the five earners. Narrative, hedging and everything-close-to-hand are among the four that rarely are.

Question 2: What is the deletion test?
  1. Delete the answer and ask again
  2. Remove the paragraph and ask whether the task could still be done correctly — correct
  3. Delete the model's memory
  4. Remove all examples

Answer: Remove the paragraph and ask whether the task could still be done correctly

If the work would still be correct without it, it was decoration - and decoration is paid for in attention.

Question 3: Why is a second-hand summary of a document dangerous?
  1. It is too short
  2. It can lose the exception that mattered, and the model then answers confidently about a world that does not exist — correct
  3. It costs more tokens
  4. It cannot be quoted

Answer: It can lose the exception that mattered, and the model then answers confidently about a world that does not exist

Half-remembered detail is worse than absent detail, because it looks like evidence.

← 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.