Chapter 9 · Talking to models: prompting and context
Context engineering
Good results come from putting the right information, tools and instructions into the context at the right time.
Chapter 9 · Talking to models: prompting and context
Good results come from putting the right information, tools and instructions into the context at the right time.
Everything is piled in. Budget full, and the model must dig for the point.
The model only sees what you put on its desk. More papers is not better; the right papers are.
You met the context window. Whatever is not inside it does not exist for the model: instructions, facts, earlier chat, tool notes.
Choose snippets for the task. The budget limits how many tokens fit. Aim for ready: the needed facts, no clutter, no stale notes.
Task: A customer asks: “When will my order 4521 arrive?”
Toy answer: I do not have enough information, so I would have to guess.
What does the model use to answer?
Why not put every note into the context?
An old note disagrees with a new fact. What is the risk?