AI Guides › Claude Mastery

The Context Window Is A Budget, Not A Wall

By Nigel Guy · 2 min read

People talk about the context window like a wall you hit — as if the conversation just stops working past some invisible line. That's the wrong mental model, and it's why people either panic about running out of room too early, or dump in far more than they need because "there's plenty of space."

The mechanism

1. Know roughly what you're working with. Current-generation Claude models run with a very large context window — Sonnet 5, for instance, ships with a 1-million-token window as both its default and its ceiling. That's genuinely large. It is not infinite, and it is not free to fill.
2. Understand that "tokens" aren't quite "words." Anthropic's newer tokenizer (used from Sonnet 5 onward) splits text into more tokens than older models did for the same amount of prose — roughly 30% more. In practice: a 1-million-token window today holds noticeably less actual writing than a 1-million-token window did a generation ago. Don't assume an old rule of thumb about "how many pages fit" still holds.
3. Spend early tokens on structure, not raw material. A short, well-organised brief early in the conversation is worth more than a wall of loosely related documents pasted in "just in case."
4. Cut, don't just add. If a document is only partly relevant, paste the relevant section and summarise the rest in your own words. You're the editor; use that role.
5. Start a new conversation when the topic genuinely changes, not because you're worried about running out of room. A fresh, focused conversation almost always out-performs one enormous thread trying to cover five unrelated things.

What to skip

Skip pasting in everything you own "for context" — a 200-page document dumped in whole crowds out the parts that actually matter to your question. Skip worrying about the exact token count for ordinary work; it only matters once you're routinely working with very long documents or very long sessions.

Guardrails

• Exact context-window sizes and pricing change with every model release — check the current model's own documentation before planning anything that depends on the precise number, rather than trusting a figure from memory (including this guide's, a few months from now).
• A large context window is not the same as good recall of everything in it. For anything that matters, re-state the key fact near where you need it used, rather than trusting it was perfectly retained from ten thousand words earlier.
• More context is not automatically a better answer. A tightly scoped question with the right three facts usually beats a loosely scoped one with thirty.

All 751 AI guides · JulieMango plans from £17/mo