Everything the Model Can See
The context window is everything the model can look at in one pass — and it is finite. So what you place inside it is a design decision, not an afterthought. New to how models read tokens? Revisit GenAI‑101 for the basics.
It All Shares One Budget
Everything the model reads and writes competes for the same finite space:
- System instructions — the standing rules for the model
- Your prompt — the request for this turn
- Retrieved documents — whatever you fetched to help
- Conversation history — the turns so far
- The output being generated — the answer takes room too
More Context Can Help or Hurt
Context is not free upside. Relevant context improves answers; irrelevant context wastes budget and can distract the model from what matters.
What to Keep, What to Fetch
Not everything belongs in the prompt all the time. Split by how stable and how often needed the information is:
- Include in the prompt — stable, always-needed info the model must have every turn
- Fetch on demand — the rest, pulled in only when the task needs it
- Coming in Module 6 — retrieval-augmented generation is how you fetch that rest well
Longer Isn’t Free
A bigger context has real costs: it generally costs more and can be slower to process. And on very long inputs, models can lose track of the middle — the ends get more attention than the center.
Build It
How to implement: pick one feature. List what must be in the prompt versus what can be fetched on demand or dropped entirely. Then cut the window down to the essentials.
- Weekly AI Tasks tracker — the parse prompt needs the task schema plus the message, not your whole history; keep the window tight.
- Personal brand site — feed each section only the relevant source material, not the entire résumé at once.
What you learned
The context window is finite and shared by instructions, prompt, retrieved docs, history, and output. Relevant context helps; irrelevant context costs money, time, and focus. Include the stable essentials, fetch the rest, and keep the window tight.