Every message you send, Claude rereads the entire conversation from the beginning.
💬 Message 1 = 500 tokens.
💬 Message 30 = 15,000 tokens
Message thirty is really messages one through thirty. Claude does not pick up where it left off. It reads everything above your new message, then reads your new message, then answers.
Anthropic's developer documentation describes each turn as containing "all previous conversation history plus the current user message," and says "previous turns are preserved completely."
One developer analyzed a 100-message Claude session. 98.5% of the tokens were spent rereading old chat history.
A long chat feels safe because nothing got deleted. It also makes Claude worse at remembering. This is the cost almost nobody watches for.
Anthropic’s engineering team describes models as working from an “attention budget.” In their words, “as the number of tokens in the context window increases, the model’s ability to accurately recall information from that context decreases.”
This is called context rot.
Nobody keeps a bloated chat alive because they enjoy the cost. They keep it alive because starting over feels like losing everything.
It is not. Ask for a handoff note before you leave by pasting this prompt.
This chat is getting long. Write me a handoff note under 200 words that I can paste as the first message of a new chat. Include what we are building, the decisions we already locked, the current draft or state, and what comes next. Skip everything we tried and rejected unless the rejection still matters.
Paste that note into a new chat and you are back at work in one message. The new chat starts at roughly 270 tokens of context instead of 15,000.
Work down the list and take the ones that fit how you actually use Claude.
One chat, one job. When the topic changes, the chat changes. A pricing question does not belong in the chat where you are drafting a job post.
Edit your message instead of arguing with the answer. When Claude misses, scroll up, tap or hover the message that caused the miss, and use the edit icon. A correction underneath adds weight, while an edit replaces it.
Abandon after three bad turns and start fresh. Three failed attempts to steer a chat back means the chat is the problem, not your phrasing.
Give research its own chat. Research pulls in large amounts of page text, and all of it rides along in every message afterward. Get your findings, write a handoff note, then draft somewhere else.
Front-load the first message so you need fewer rounds. Most long chats are long because rounds two through eight were clarifications you could have supplied at the start.
Ask for the changed part, not the whole redraft. Twelve rounds of full redrafts leaves twelve full drafts in the chat.
Batch your questions into one message. Five separate questions mean five separate reloads. One message with five numbered questions means one.
I recommend that you also read the post below.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.