The Context Window X-Ray — robert@barcik.training

Most people assume the text they type is all the LLM sees. It isn't — not even close. Modern AI apps engineer the context window: on every single call they assemble a fresh payload of system instructions, retrieved memories, skill catalogs, tool definitions, uploaded documents, search results, and conversation history — and your message is usually the smallest part of it. Follow Maya, a customer-operations lead, through six messages to her company's AI assistant. Left: the chat she sees. Right: the payload the model sees. Click any block to read its full contents.

What Maya sees the illusion

✨

Acme Assistant

Your company AI · always here to help

📎 refund-policy.pdf

What's Happening

Maya's first message is staged in the input box. Press Send and watch the right side.

The chat window looks like a clean two-way conversation. The right pane shows what is really shipped to the model each time.

What the model sees the reality

Context Window 0 / 16,000 tokens

Payload of this API call — click blocks to inspect

Nothing has been sent yet.
The window is assembled fresh for every call.

API calls made

Maya typed, in total

0 tok

Sent to model, in total

0 tok

the whole window is re-sent every call

Takeaway

The model is stateless. The app is the author.

Every call starts from zero — "memory" means the app re-packs everything it wants the model to know, every single time.
Across six turns Maya typed a few hundred tokens; the app shipped tens of thousands. Most of the prompt is text she never wrote and never sees.
Context is a budget: memories are retrieved selectively (and flagged as possibly stale), skills load lazily, documents get truncated, and old turns are silently compacted away. Deciding what enters the window, and at what fidelity, is context engineering — it is most of what makes a GenAI app good.

🔁 Try it again on an 8K window (top right) — watch how much the app has to sacrifice, and how early. Then compare with 200K: nothing is lost, but every call re-sends ~15,000 tokens.