info snack
Context windows
The amount of material a model can consider at once.
companion card
A context window is the amount of prompt, history, and retrieved material a model can consider at once.
What it means
Instructions, chat history, retrieved documents, examples, tool results, and user input all consume context. Output tokens need room too.
Why product teams care
A huge context window can hide the answer in noise. Structure, ordering, summaries, and retrieval quality often matter more than raw length.
Understudy angle
Understudy can compare context strategies with evals: shorter prompts, retrieved context, examples, schemas, or specialist routes.
take this with you
More context helps only when the right information is present and easy to use.
Remove irrelevant context from one failing prompt and rerun the same eval.