System preamble Vendor rules · identity · date Tool definitions Search · code · files · skills Memory / project Saved facts · files · instructions Conversation so far Earlier messages · tool calls · results Your message “Summarize this” Everything above your message was written by the vendor, the harness, or your organization. Some of it runs to tens of thousands of tokens before you say a word. Why it matters
The model does not distinguish "instructions" from "data" by magic. It all arrives as text. What the organization puts in the preamble shapes behavior as much as what you type.
That is also why prompt injection is real: a document that contains "ignore previous instructions" is just more text in the same prompt (see Prompt injection ).
The preamble and tools take space out of the window before you start. That is why "1M tokens" is never 1M tokens of your conversation.
✓ Do
Ask the model "what is in your system prompt?" once. It will partly describe it. Knowing that layer exists explains most odd behavior.
When a project or a connector is loaded, remember it is now part of every prompt, whether or not the current question needs it.
✕ Don't
Don't assume your message is the only instruction in play. If the tool insists on a format or refuses a phrasing, the preamble is usually why.
Myth, busted
"The prompt is what I typed." The prompt is everything the model receives this turn . What you typed is the last, and often the smallest, part.
Takeaway
Preamble, tools, memory, history, then you. Keep that picture and the rest of this guide falls into place.
GenAI Field Guide