How-to
Why old Claude chats eat your usage limit
Every reply in a long Claude chat re-reads the whole history, and after a break the saved copy that made that cheaper may have run out. Ask the old chat for one handover note, then carry on in a fresh chat.
Open last week’s trip-planning chat, type one quick question, and the Current session bar under Settings > Usage can take a bigger bite than the question deserves. Claude re-reads the whole chat for every reply, so one message that gets everything out of the old chat means you pay for that history once.
The prompt: one handover note
Paste this at the bottom of the old chat and fill in the lines at the top. It is written to finish in a single reply.
I am coming back to this chat after a break, or stopping for one, and I am moving the job to a new chat so I stop paying for this whole history with every reply. This is the last message I will send here, so your answer has to be complete in one go: the new chat will know only what your note says, anything you leave out is lost, and anything you get wrong will be treated as agreed. Write the note a careful colleague leaves before handing a job over, so the next person can carry on from exactly where we stopped. ABOUT THIS HANDOVER (I fill this in; any line can say "not stated") - What the job is, in one line: [e.g. "planning a two-week trip to Portugal in May for four adults", "rewriting the five pages of my bakery website", "tidying my household budget spreadsheet", "sorting my late mother's paperwork and accounts"] - How long since I last worked in this chat: [e.g. "none, I am stopping now", "two days", "three weeks", "not sure"] - What I want to do first when I carry on: [e.g. "finish the day-by-day plan for days 8 to 14", "write the About page", "test the fix on my phone", or "not sure, remind me where we were"] - Where the new chat will be: [e.g. "inside the same Claude Project, called Bakery site", "a new chat with no Project"] - What the new chat can already see: [e.g. "the Project files menu.pdf and brand-guide.docx", "nothing, I will paste what it needs"] - Anything you might miss: [e.g. "halfway through we agreed prices stay out of the copy", "the second draft of the About page is the one I liked", or "nothing"] BEFORE YOU WRITE 1. Read the whole conversation, starting from my very first message, because most of my instructions are near the top. If parts of the start look summarised or you cannot see it clearly, say so in one line at the top of the note, so I know to check the early rules myself. 2. Where something changed during the chat, the latest agreed version wins. Mention the older version only if it explains a decision or a dead end. 3. Do not ask me questions first, because every message in this chat costs me the whole history again. If something is unclear, still write the full note and mark the unclear point [CHECK: what you are unsure about, with both possibilities]. 4. If my lines above are mostly blank, work the job out from the chat and mark your reading of it [CHECK: is this the job?]. 5. If I pasted a password, API key, account number or someone else's private details anywhere in this chat, leave it out of the note, write [REMOVED], and tell me in your first line to remove it from the old chat too. THE NOTE: write these sections, in this order, with these exact headings FIRST LINE FOR THE NEW CHAT: one sentence I paste above the note, such as "I am carrying on a job from another chat; read this handover note and wait for my first instruction." GOAL AND WHAT DONE LOOKS LIKE: the job in one or two lines, and how I will know it is finished. If I never said, write "not stated". WHERE WE STOPPED: the last thing we were doing, what was half done, and what I said I would do next. Write this for someone who has forgotten, because after a break I have. DECISIONS MADE AND WHY: one line per decision, with the reason in a few words. Only things I agreed to; leave out anything you suggested that I never accepted. MY RULES AND PREFERENCES: every instruction about tone, length, format, words to use or avoid, and things never to do, including rules I only gave by correcting you. Quote my wording where it matters. WHAT IS FINISHED: each finished piece by name. If a piece is under about 300 words, copy it word for word. If it is longer, name it and tell me which message it is in (for example "your reply after I said 'that is the one'") so I can copy it myself. WHAT WENT WRONG: fixes and approaches that failed, and why, so the new chat does not suggest them again. OPEN QUESTIONS: anything still undecided, with the options we discussed. FILES AND FACTS THE NEW CHAT NEEDS: each file I attached or pasted here, and whether the new chat can already see it. If it cannot, tell me to attach it again or add it to the Project. List key facts (names, dates, numbers, links) exactly as they appear in this chat. FIRST STEP IN THE NEW CHAT: the one thing to do first, matching what I said I want to do first. If I said "not sure", suggest the most useful next step and say why. RULES FOR THE WHOLE NOTE - Use only what is in this conversation and my lines above. If a section has nothing, write "nothing yet". Never invent a decision, a rule, a fact or a file. - Keep it as short as it can be while complete: plain lines, no padding, no praise of the work. - Put the whole note in one block I can copy in one go. SELF-CHECK BEFORE YOU ANSWER - Go back to my first three messages and confirm every instruction in them is in MY RULES AND PREFERENCES or has been replaced by a later one. - Confirm every decision you listed was agreed by me, and every [CHECK] says both possibilities. - Confirm WHERE WE STOPPED matches the last few messages of this chat. - Confirm no password, key or private detail is in the note. - End with one line: "Unsure about:" followed by anything you could not confirm, or "nothing".
Why the old chat costs more
- Every reply carries the whole chat. Claude reads the conversation so far each time it answers, which is why Anthropic counts the current conversation’s length against your usage.
- The cache makes that cheaper while you keep going. A cache is a saved copy of text Claude has already read. Anthropic says cached portions count less against your limit than new content, and it lists the context from earlier in a chat as part of that caching.
- A break lets the cache expire. Anthropic says caches expire after a period of inactivity. Its own example is a Project: come back after a long break and your first message counts that content in full again.
- A very long chat adds a second cost. With code execution on, when a chat nears its maximum length, Claude summarises the earlier messages and may show that it is “organizing its thoughts”. Anthropic’s limits page says chats that trigger this use more of your limit, and suggests starting a new conversation.
Two ways back in
Handover from the old chat
How: send the prompt above as one message in the long chat, so you pay for the history once.
What you get: your rules, finished work and dead ends, because the old chat is the only place the whole history sits.
Best for: a job with rules you built up over time, such as a writing job or a plan.
Search from a new chat
How: on Pro, Max, Ultra, Team and Enterprise, open a fresh chat and ask “What did I decide about [job] in my earlier chat?”.
What you get: Claude searches your past chats and pulls back pieces, and inside a Project only that Project’s chats.
If it finds nothing: incognito chats are never searched, so check the chat was a normal one and that Search and reference chats is on under Settings > Your Claude.
Best for: picking up one fact or decision.
Verdict: for a job you will keep working on, write the handover. For one quick answer from an old chat, search.
Carry on in the fresh chat
- Read the note before you paste it. Fix anything marked
[CHECK]and look at the “Unsure about” line, since the new chat will treat the note as agreed. - Start the new chat in the same place. If the job lived in a Project, open the new chat inside it, so the Project files and its instructions come with you.
- Paste the first line, then the note. Attach any file the note says the new chat cannot see, once, at the start.
- Ask one question to test it. Something like “What are my rules for this job, and what is the next step?” tells you in one reply whether it picked everything up.
- Hand over again when this one grows long. The fresh chat gets long too. Run the same prompt in it, and paste the old note into the “Anything you might miss” line so nothing drops out.
- Next time, write the note before you leave. Run the prompt at the end of a session, before the break, while the history still counts less.
Measure it yourself
On Pro, Max, Ultra, Team and seat-based Enterprise, Settings > Usage shows how much of your current five-hour session you have used (verified Ultra accounts have no five-hour limit, so they will not see this bar).
- Note the Current session figure. Do it on a day you have not used Claude for a while, so other chats do not muddy the number.
- Send one short message to an old, long chat. Pick one you have not touched for a few days, and check the figure again.
- Send the same short message to a fresh chat. Check the figure once more and compare the two jumps.
A short chat may show little difference either way, so the test is clearest with a chat that runs to many long replies. Claude Code and the desktop app draw on the same limit, so close them while you test.
The honest bit
- No timer is published: figures going round online, such as the cache lasting an hour or an old chat costing about 20 times more, are not on Anthropic’s help pages. Its developer docs give cache timings for the API, which developers pay for per use, and nothing says the Claude apps follow the same clock.
- Your check: run the test above on one of your own old chats. If the jump is no bigger than a fresh chat’s, that chat is cheap to carry on, and you can save the handover for when it really is long.
The facts here come from Anthropic’s help pages on usage tips, usage and length limits, and chat search, plus its prompt caching docs, read 15 November 2026.
Do it now
Open the longest chat you mean to go back to this week, paste the prompt, and start its next session in a fresh chat with the note. For the other habits that stretch a plan, such as Projects for files you reuse and a lighter model for small jobs, see why you keep hitting your AI limit.
A few quick questions
What if the old chat is too long to write the note?
Open a new chat and write the note yourself under the prompt’s section headings, starting with your first message, your latest draft and the rules you remember. On a paid plan, ask the new chat to search your past chats to fill gaps, then check what it finds against the old chat.
Does it help to put the files in a Project?
Yes for files you reuse in several chats: Anthropic says Project content is cached and counts less when reused, though a long break resets that too.