Long inputs and what gets used

Context windows have grown enormously. It is tempting to read that as “I can paste everything now”, and for many tasks that produces worse results than pasting less.

Capacity is not attention

A large window means a large amount of text can be supplied. It does not mean every part is weighted equally when producing the answer.

The well-documented pattern is that material at the beginning and end of a long input is used more reliably than material in the middle. A fact buried halfway through a long document is measurably more likely to be missed than the same fact near either edge.

This has been demonstrated repeatedly across models, and while each generation improves, the shape persists. Design around it rather than hoping.

Put the instruction where it will be read

Practical consequence, and the highest-value thing in this lesson.

With a long document, do not lead with a short instruction and follow with forty pages. The instruction is then a small thing a long way from where the answer gets produced.

Better:

[document]

---

Using only the document above: [instruction]

Or, for anything important, put it in both places — a brief framing before so the material is read with purpose, and the full instruction after so it is the last thing in the context.

The repetition feels inelegant and measurably helps.

Less material, better answers

The counter-intuitive one.

Given a document and a question about section 4, supplying only section 4 and its surrounding context usually beats supplying the whole document. Relevant material is not diluted, and there is less opportunity to draw on the wrong part.

So before pasting everything, ask what the task actually needs. Often it is one section, one table, one thread — and cutting is faster than the rerun you would otherwise do.

This also applies inside a chain (L4): pass forward only what the next step requires.

When you must supply a lot

Sometimes the whole thing genuinely is the input. Three things that help.

Chunk and combine. Process in sections, then combine the results in a second pass. Slower, more reliable, and each chunk is small enough to be read properly.

Label the parts. Give each section a clear marker — [SECTION 3: PRICING] — so the model can locate material and you can ask about it by name.

Ask for location. “Quote the sentence you used and give its section.” This is E4·L3’s traceability rule and it doubles as a check: an answer that cannot point at where it came from often did not come from there.

Watch the cost

Long inputs cost more per call, and in a chain they cost more at every step that carries them.

Trimming the input is usually the largest single cost reduction available in an automated workflow, and it typically improves quality at the same time. That combination is rare enough to be worth checking for deliberately.

Try it now (6 minutes)

Take a long document and a specific question about it. Ask twice: once with the instruction first, once with the document first and the instruction last.

Then ask a third time with only the relevant section. Compare all three — most people are surprised by the third.

Check your understanding

1. In a long input, material is used least reliably when it sits: At the beginning In the middle At the end

2. With a forty-page document, the instruction is best placed: Before the document, kept brief After the document, or in both places In the system prompt only

3. Supplying only the relevant section rather than the whole document: Usually produces worse answers due to missing context Usually produces better answers and costs less Makes no measurable difference

Recap

A big context window supplies text; it does not guarantee even attention, and the middle of a long input is the weak spot. Put your instruction after the material, or in both places, and prefer supplying only what the task needs. Where the whole thing really is the input, chunk it, label the sections, and require answers to cite where they came from.