What reaches the model
The decision depends on this specific input, not everything held by the application.
Links containing ?t= open the video at a specific second.
The ideas to retain
First, what does “context” mean here?
The word context is overloaded. In this chapter, it means the effective input available to one specific inference.
Prompt engineering is one part of the problem
Prompt engineering still matters. Ambiguous, contradictory, or excessively prescriptive instructions can degrade behavior even when the rest of the system is sound.
Context is a snapshot, not the full application state
This distinction prevents a particularly dangerous mistake in agent systems: treating “the system knows X” as equivalent to “the model can use X now.”
Jump directly to a section
Read the reviewed transcript
This video has no narration. This transcript reproduces its on-screen text; it does not invent a spoken track.
Available is not visible
The application holds documents, memory and tools. The model does not automatically receive everything.
The assembler selects three items for this inference. Everything else stays outside.
The decision depends on this specific input, not everything held by the application.
An outdated policy
The prompt asks for the current policy on a twenty-day-old order.
Copy A allows thirty days. Rewording the prompt does not update that source.
Current revision B allows fourteen days. With the same prompt, the evidence requires a different decision.
Filter before ranking
Three passages appear relevant, but one belongs to another customer.
The access boundary excludes it before utility is compared. Similarity grants no permission.
Only eligible items can compete for space in the context.
The next turn changes
Context C0 contains a hypothesis about the failure. The test has not returned yet.
The actual result contradicts that hypothesis. This observation did not exist in C0.
The assembler puts the result in C1 and removes the invalidated hypothesis.
Load when needed
Stable instructions enter at the start. Large files remain behind references.
A new question triggers a read of the relevant file, not the whole repository.
On-demand loading saves context but adds a read that can fail.
Reconstruct the turn
An answer alone does not reveal which information the model received.
The receipt keeps IDs, revisions and reasons for inclusion or exclusion.
This separates missing evidence from evidence that was received but misused.
