Semantic Machines explainers · Dominik Lukeš

See what the model actually receives

  • A language model has no memory between requests.
  • On every turn, the software around it, called the harness, sends everything again: the instructions, the tools, the chat so far and anything it has read.
  • It is not a one-line instruction. A modern agent starts with 15,000 to 40,000 tokens before you type: 30 to 80 printed pages, almost a little book.
  • The window has a fixed size, about 400 pages for 200,000 tokens, so the pile can fill up.
  • Step through a scripted example, then open Show context to see each part, its size and what is new.
Context window200,000 tokens · about 400 pages
One request. All of this is sent again on the next turn, plus whatever was added.

Scenarios

Eight scripted examples. The first shows the idea in miniature; the next two use real, published instructions and tools; the rest show agents doing everyday work. Pick one, then step through it at your own pace.

Example agent Context Explorer

Embed this scenario

Paste this into a web page or slide tool that accepts iframes. It shows only this chat, its step bar and the context panel.

Link:

Token counts are approximate: the real sources were counted with OpenAI's cl100k tokenizer, which is not the native tokenizer of every model. Where a block is long, the text shown is an excerpt and the count is for the full block. Each block says whether it is real, a real format with illustrative content, an estimate or illustrative.

How it works

Open Show context in any example. The panel lists everything the model received for that step, as coloured blocks. Each colour is a kind of block.

  1. The bar

    Shows how much of the window is in use. Zoom it to see the parts more clearly when little is used.

  2. The New tag

    Marks blocks that were not in the previous request. Everything else was sent again.

  3. Follows the step

    While the panel is open, Back and Next update it, so you can watch the context grow.

  4. Tokens and pages

    A token is a word or part of a word. As a rule of thumb, 500 tokens fill one printed page, about 375 words. A 200,000-token window holds about 400 pages. Some models offer 1,000,000 tokens, about 2,000 pages.

  5. Real or estimated

    Real text is quoted from a published source. Real format uses a real layout with invented content. Estimate gives a size worked out from the sources. Illustrative is made up to show the idea.

  6. Compaction

    When the context nears the limit, the harness has the model summarise the work so far, then starts a new, much smaller context from the set-up, the summary and a few recent items.

The blocks

About

Context Explorer is one of the Semantic Machines explainers, made by Dominik Lukeš.

All conversations are scripted. No model is running, and the files, websites and figures in the chats are invented.

The system prompts and tool definitions marked Real are quoted from published sources, shortened with a note of what was left out:

  • Anthropic, Claude Fable 5.2 system prompt (published 9 November 2026). It covers behaviour only; Anthropic does not publish the app's tool definitions.
  • OpenAI, Codex CLI at commit c0d4726: base instructions, tool specs, skills prompt, environment format, compaction prompt and compaction limit. Apache License 2.0.
  • Google, Gemini CLI at commit 4fa64e7: tool schemas, used only to work out the average size of a tool definition. Apache License 2.0.

Sizes marked Estimate use an average of about 375 tokens per tool definition, taken from 22 real definitions in those two open-source tools. Real harnesses differ in detail, but the pattern is the same: the model sees one long piece of text, rebuilt on every turn.