Quick story so the rest of this page makes sense. Every time you send Claude a message, it doesn't just read your new message — it re-reads the entire conversation from the beginning. Every previous question, every previous answer, every file you pasted in.

So a fresh chat with one short question = cheap. The same short question 30 messages deep into a long chat = wildly more expensive, because Claude is re-processing everything that came before it. People have measured this and found that more than 90% of the "tokens" they burn aren't on their actual question — they're on Claude re-reading old history that nobody needs anymore.

Once you understand that, the playbook becomes obvious: don't let your chats get long. That's the whole game. Everything below is just five different ways to stop that from happening.

This is the move that changes everything. Once you've done it a couple of times, you'll wonder why nobody told you sooner.

The pattern: Every 15-20 messages — or any time a chat starts to feel slow — you ask Claude to summarise what you've worked on so far. Then you start a brand new chat and paste that summary in as the very first message.

That's it. That's the whole tactic. You've kept the bits that matter (decisions made, files created, context) and ditched the bits that don't (every back-and-forth, every "wait actually" course-correct, every tangent). New chat, fresh tokens, way faster Claude.

The exact prompt — copy this

Paste this at the end of your old chat. Claude will give you a tidy handoff doc. Then start a new chat and paste the doc in as message #1.

I want to start a fresh chat to save tokens. Please write me a handoff document I can paste into a new conversation, so the next chat picks up exactly where this one leaves off. Include:

1. The goal — what we're actually trying to do
2. What's already been decided (and any open questions)
3. Any files, links, or specific details the next chat will need
4. The very next step

Write it as if you're handing off to a teammate who has no context. Skip the chit-chat. No preamble.

When to do it

• Every 15-20 messages — set yourself a mental timer. The longer you wait, the more painful the chat gets.
• The moment Claude starts repeating itself or feels sluggish — that's the signal it's working too hard to remember.
• Before any big task switch — finishing one piece of work and starting another? Handoff first, every time.
• At the end of a working session — so tomorrow's chat starts strong instead of wading through yesterday's mess.

Once the Handoff is in your toolkit, these four are the rest of the kit. None of them require any technical knowledge — they're just habits you build over the next week.

Here's the thing nobody mentions: Claude's replies cost about 5x more than your messages. Input tokens (your stuff) are cheap. Output tokens (Claude's stuff) are the expensive ones. So the single fastest way to save tokens isn't even about your prompts — it's about telling Claude to shut up faster.

When you ask a question, Claude's default is to be polite, restate your question, walk you through its reasoning, summarise its answer, and offer follow-up suggestions. That's a lot of words you didn't ask for. Each one costs you.

The fix is one line at the top of any prompt:

Skip the preamble. Don't restate my question. Give me the answer only.

You can also set this once-and-forget by going to Claude's settings and using a custom style or profile preference that says "Be concise. Skip pleasantries. No preambles or summaries unless I ask." Claude follows it across every new chat without you typing it again.

If you read three Reddit threads on saving tokens, you'll see a hundred "tips" that don't actually matter for non-technical users. Here's the list of stuff to not waste your time on:

• Prompt caching. You'll see this everywhere. It's a developer/API thing — Anthropic does it automatically when it makes sense, you don't have to set it up. If you're using Claude.ai or Claude Code as a normal human, ignore it completely.
• The exact token counts of every prompt. Nobody's actually tracking their tokens to four decimal places. The five habits on this page do 95% of the work — chasing the last 5% is a rabbit hole.
• Special "token-saving" prompt formats. You don't need to write everything in markdown tables or learn arcane prompt syntax. Just write clearly. Long rambly prompts are wasteful, but a normal-length English question is totally fine.
• "Tricking" Claude into being faster. Some posts swear that asking nicely or using emoji or saying "please be brief" works — and to be fair, "be brief" does help. But anything more elaborate than that is folklore.
• Worrying about every image you upload. Yes, image attachments cost more tokens than text. No, you don't need to never use them. Just don't paste 12 screenshots into one chat when one would do. Crop tighter when you can.

If you want to be in a room with other women figuring this out — asking questions, sharing wins, getting unstuck together — come into the Wright Mode Membership. We share workflows, prompts, and skills like the ones in this playbook every week.