---
title: The context window — Around the model, lesson 1
description: Context is everything a model reads in one call: instructions, tools, pages, the conversation, your question. Where each part sits, why the middle counts least, and what context engineering decides.
url: https://instibrain.ai/guides/around-the-model/the-context-window
updated: 2026-11-08
site: https://instibrain.ai/llms.txt
---

lesson 1 of 9 · around the model

# The context window: everything the model reads in one call.

Context is everything the model reads in one call, not only your question. It is all the model knows about you and your company.

## new words

**Context**
Everything the model reads in one call: the instructions, the tools, the pages read, the conversation, and your question. The field also calls it the prompt; the line you type is only its end.

**Lost in the middle**
The model attends most to the start and the end of the context, and least to the middle.

**Context rot**
Answers get worse as the context window fills, before it is full.

**Context engineering**
Deciding what goes into the context window, in what order, and what stays out.

**Progressive disclosure**
Loading in steps: who the AI is first, then an index of what exists, then one page at a time.

## the lesson

The context window is the space; the context is what fills it. AI basics, lesson 1 gives its size in tokens. In one call it holds, in this order:

- the instructions: the app's own, and your instruction files;
- the tools the app offers, each with its name and use;
- the pages and files read for this task;
- the conversation so far, with each tool's result;
- your question, last.

The weights know language and the world. Everything about your company and this conversation is in the context, or nowhere. The model keeps nothing between calls, so every message sends the whole context again (AI hygiene, lesson 1).

Where a line sits matters. The start and the end get the most attention, and the middle the least: lost in the middle. So the instructions go first, and the question goes last.

A full context window gets worse before it is full. Old results and wrong turns stay in it and keep pulling on the answer: context rot. When it is nearly full, the app replaces the old conversation with a summary: compaction (Claude basics, lesson 2).

So every call is a choice of what goes in: context engineering. An example company keeps 2,000 pages of about 500 tokens: 2,000 × 500 = 1,000,000 tokens, a whole context window before the first question. Loaded in steps, the AI reads who it is, an example 5,000 tokens; then one folder's index, a line for each of its 100 pages at about 20 tokens, 2,000 tokens; then the two pages the task needs, 1,000 tokens. That is 8,000 tokens, 0.8% of the context window, and every one of them is about the task.

One call's context, from the top. All five parts are sent again with every message.

## try it

In a Claude Code window, type this and press Enter:

```
/context
```

Then ask Claude to read one long file in the folder, and type `/context` again.

## you should see

- A grid, then a list of what uses the context window: the instructions and the tools first, the messages after.
- The second time, the messages part is larger by about the file's size in tokens. The file stays in the context for the rest of the session.

---

instibrain is made by WebWork AI, Inc. Everyone uses the same models · the difference is what yours knows.
