Keeping a long chat useful — and knowing when to start fresh

Working with Claude — CC BY 4.0

A conversation with Claude is not like a phone call, where each of you remembers the gist and moves on. It is more like a shared desk. Everything you have both written in this chat stays on the desk, and Claude reads the whole desk again before every single reply. That is why a fresh chat feels sharp and a very long one starts to feel foggy, slow, or forgetful. Once you understand the desk, you can keep chats useful for as long as you need them — and spot the moment when starting over is the better call.

The context window, in plain terms

The “context window” is the size of that desk: the total amount of text Claude can hold in mind at once, counting your messages, its replies, and anything you have attached. It is measured in tokens — roughly, chunks of words. You do not need to count tokens. You just need the intuition: there is a limit, everything in the chat competes for the same space, and the whole lot is re-read each turn.

Current windows are large — enough to hold a long document or a book-length conversation — and the exact size depends on the model and plan you are on. The figures move as Anthropic ships new models, so treat any specific number you read online with caution and confirm the current one in-app or in Anthropic’s help centre rather than trusting a memory of it.

What happens as a chat fills up

When a conversation gets close to the limit, Claude does not usually fall over. On current models it will quietly summarise the earlier parts of the chat to make room, keeping the thread going while your full history stays available for it to refer back to. Most of the time you will not notice this happening.

Two things matter anyway. First, this tidying is a summary — compressed, not word-perfect — so fine detail from early on can lose its edge. Second, in rare cases (a huge attachment, an unusually long single message) you can still hit a hard wall where Claude tells you the conversation has reached its maximum length. If that happens, don’t fight it — start fresh with a good handoff, which we come to below.

(The exact behaviour here depends on the model and settings; confirm current details in Anthropic’s help centre.)

Iterating without cluttering the desk

Most work with Claude is a back-and-forth: ask, read, refine. A few habits keep that loop clean.

• Batch related questions. Three separate messages means Claude re-reads the whole chat three times and you burn through your usage faster. If you can ask it in one, do.
• Be specific, then correct. Say what you want, look at what you got, and tell it plainly what to change — “shorter, drop the last paragraph, keep the middle example.” Steering beats re-asking from scratch.
• Quote it back to itself. If Claude has drifted from an instruction, paste the instruction again. It responds to what is on the desk right now, not to what you meant ten messages ago.

When a fresh chat beats a long one

A long chat is not automatically a good chat. Start a new one when:

• It has drifted off-topic and half the desk is now irrelevant — the clutter drags every reply.
• You are changing subject entirely — a new task deserves a clean desk, not the residue of the last one.
• Claude keeps repeating a mistake you have already corrected. Sometimes the error is now baked into the history it re-reads, and the cleanest fix is to reset.
• Replies have slowed or gone vague as the chat has grown very long.

The trick to starting fresh without losing your work is a handoff. Ask the current chat: “Summarise everything we’ve decided and where we’re up to, so I can carry it into a new conversation.” Read the summary, check it is right, then paste it as the first message in a new chat. You keep the substance and drop the clutter — and because you read it before reusing it, you stay in control of what carries over.

Two features can help with continuity across chats: Projects, which keep related conversations and reference material together, and Claude’s memory, which can carry preferences and recurring context between chats. Both are worth exploring, both are under your control in settings, and the exact menus shift as the app updates — look under settings for the capabilities or memory area, and confirm in-app.

Keeping the judgment with you

The through-line of this whole course applies here too: you direct the conversation, you review what comes out of it, and you decide when it has earned a reset. Do not let a long chat coast on autopilot — when it starts to wander, re-state the goal, correct the drift early before it compounds, or hand off and start clean. The thread is yours to manage. Managing it well is most of what makes Claude reliable over a long piece of work rather than merely quick at the start.