In lesson 2.5 you met the context window: Claude's working memory for a single chat. We left the mechanics for later. Later is now.
This one pays off fast. Once you can picture what fills the window and why a long chat goes fuzzy, you stop fighting Claude and start steering it. The reward is simple: sharper answers, fewer do-overs, and chats that do not fall apart at the two-hour mark.
First, a token
Claude does not read letters or whole words. It reads tokens: small chunks of text. A token is about four characters, or roughly three-quarters of a word. So 100 tokens is about 75 words, and a tidy page of text is somewhere near 500 tokens.
Why care about a unit of text? Because tokens are the currency of everything. The window's size is counted in tokens. Cost is counted in tokens. Speed tracks the token count too: more tokens in, more to chew through, longer to answer. So "how many tokens is this?" is really "how much space, money, and time will it take?"
What the window actually holds
Picture your desk. The context window is the desk surface: everything Claude can look at while it works. It is not a filing cabinet of all your past chats, just what is on the desk right now, in this chat.
Three things share that surface:
- The files you added. Uploads, pasted text, project material.
- The whole conversation so far. Every message, yours and Claude's, back to the top.
- The reply being written. A slice is always kept clear for the answer, so the space you can fill is a bit smaller than the full number.
How big is the desk? On paid plans, chatting in claude.ai, it runs from 200,000 tokens up to 1,000,000, depending on the model you pick. The newest model reaches the full million; most sit at 200,000. To put a million tokens in human terms, that is roughly 750,000 words, or a couple of thick books. It sounds endless. It is not, and here is why that matters.
Why long chats drift
The desk is big, but it fills. When a chat nears the limit, Claude does something sensible: it summarises the earlier parts to make room, then keeps going. Your full history is still saved, but what Claude is actively holding becomes a shorter summary of the start plus the recent detail. (On paid plans this is automatic, as long as file creation is switched on. You might catch Claude "organising its thoughts" mid-chat; that is the tidy-up at work.)
A summary is smaller than the real thing, so fine detail can slip. That is what drift feels like from your side:
- It repeats a point it already made.
- It forgets an instruction you gave near the top.
- It gets blander, or a touch vaguer, than it was an hour ago.
- It contradicts something agreed earlier.
There is a subtler version too. Even below the limit, a very long, cluttered window makes it harder for Claude to find the part that matters, the way a crowded desk buries the one page you need. Anthropic calls this context rot: more stuff in the window is not always better. What is in it matters as much as how much fits.
When to start fresh, and how to carry work over
The fix is not a trick. It is a habit: one topic per chat, and start fresh when the topic changes or the answers start slipping. A new chat is a clean desk, the whole window back at your disposal.
The worry, of course, is losing the thread. You do not have to. Hand the work over on purpose:
Before we stop, write a short handoff I can paste into a new chat.
Cover: what we've decided, the current draft, and the open questions.
Keep it tight.
Take that summary, open a fresh chat, paste it in, and carry on. You keep the thread and drop the dead weight. For anything you return to often, put the summary in a project instead (that is lesson 4.3), so it is waiting for you next time.
Keep the window working for you
A few small moves leave more room for the real work:
- Feed it lean material. As lesson 2.5 warned, a PDF is read as images and a Word file carries hidden styling, so both weigh a lot. Plain text and Markdown carry the same words for a fraction of the space. When you control the format and the material is big or reused, prefer
.mdor.txt. - Switch off what you are not using. Web search, connectors, and extra tools are token-hungry just by being switched on. Turn off the ones this chat does not need.
- Lower the effort, or turn off extended thinking, for routine work. Both spend tokens you may not need for a quick task.
- Use projects for big reference sets. A project loads only the relevant slice of your files into the window rather than the whole pile, so you can lean on far more material without clogging the desk. More on that in lesson 4.3.
The rule of thumb: a clean, focused chat beats a giant sprawling one every time. Give Claude what the task needs, and not much more.
Where this goes next
You now hold the two levers that decide how well any long piece of work goes: what you put in the window, and when you clear it. That is the quiet skill behind every power user.
Next in Part 4 we stop managing single chats and start reusing your best work. Claude Skills package a set of instructions so Claude does a job the same way every time. Projects keep files, context, and chats together. Both build straight on what you just learned.
Try it
Open a long chat you already have, or a fresh one you push for a while on a real task, until you can feel it start to slip: repeating itself, forgetting an early instruction, or going vague.
Now test the handoff. Paste in the summary prompt above, take what Claude gives you, and start a new chat with it. Ask your next question there. Notice how the clean window answers sharper than the tired one did. That gap is the whole lesson: same task, better result, just because you cleared the desk.
(General exercise. Role and industry versions come once accounts are in.)