How to save tokens in Claude and ChatGPT: 18 habits that cut AI usage without cutting quality

Every AI plan runs on tokens. When your team hits a usage limit halfway through the day, or the API bill jumps, it's almost never because the work was hard. It's because the way people use the tool burns tokens they never needed to spend.

This guide explains what tokens are, where they actually go, and the 18 habits we use and set up for clients to get the same or better work out of Claude and ChatGPT for a fraction of the usage.

What is a token?

A token is a small chunk of text. AI models don't read words, they read tokens. A short, common word is usually one token, while a long or unusual word gets split into several pieces. As a rough guide, 100 tokens is about 75 English words.

Everything the model reads counts, and so does everything it writes:

• Your message, plus every earlier message and reply in the chat
• Files, screenshots and pasted text
• Instructions sitting behind the chat (project instructions, memory, custom settings)
• Connected apps and tools, whose descriptions the model has to read
• The reply itself, including any hidden "thinking" the model does first

On Claude's API, output tokens cost five times as much as input tokens, so what the model writes back matters even more than what you send it.

Where tokens really go

The single biggest cost most people miss: every time you send a message, the model re-reads the entire conversation so far. Message one is cheap. Message forty, in a chat that has covered five different jobs, drags all forty messages, every file and every reply along with it.

So the savings come from three places: keeping what the model re-reads small, getting to the result in fewer messages, and asking for only the output you need.

Chats and projects

1. Compact long threads before going back in. Before you return to a long conversation, compact it (in Claude, type /compact where your app supports it) or ask for a short handover summary and paste it into a fresh chat. The history gets swapped for a summary, so every later message costs less.
2. Work inside projects. Claude Projects and ChatGPT Projects keep the context, memory and instructions for one area of the business in one place. You stop re-explaining the background at the start of every chat.
3. One task per chat. When a job is done, start a new chat for the next one. A chat covering five jobs makes every reply pay for all five.
4. Keep project instructions and memory lean. These load into every chat in the project, so every extra line is paid for again and again. Keep them to rules that apply to most chats and review them monthly.
5. Edit your message instead of correcting it. If the model misread you, edit your original message and resend. A follow-up correction leaves the wrong answer in the chat, and it gets re-read with every message after it.

How you ask

6. Ask for the outcome, not lots of questions. Ten small back-and-forth questions means ten full re-reads. Describe what "done" looks like ("build X, test it, give me a summary table") and let the model make the small calls.
7. Borrow how the experts do it. Find the best people in your field and how they approach the task, then ask the model to work their way. A proven method on the first try beats trial and error across twenty messages.
8. Write in plain, short words. Tokens are built from pieces of words, so one long or unusual word can cost more than three or four short, everyday ones. Cut the filler ("I was wondering if you could possibly...") and say what you need.
9. Put everything in one message. Goal, context, an example and the format you want, all at once. If you have several small requests, send them as one numbered list.
10. Ask for short answers. Because output costs more than input, a 900-word answer you skim is the most expensive waste of all. Ask for "3 bullets", "one paragraph" or "just the table".

What the model reads

11. Point to exactly what it needs. "Have a look through everything" sends the model through files, pages and inboxes. Name the exact document, page, tab or email instead, and paste only the section that matters.
12. Crop screenshots and trim big files. Images and long files are read in full. Crop screenshots to the part you're asking about, and tell the model which pages or columns matter in a long PDF or spreadsheet.
13. Switch off connectors you don't need. Each connected app adds tool descriptions to read. Turn off the ones a chat won't touch.
14. Turn repeat instructions into skills. Claude skills (and custom GPTs or saved instructions in ChatGPT) only load their full instructions when they're needed, and you stop retyping the same steps.

Bigger jobs

15. Save deep research for big questions. Research modes, extended thinking and long chains of web searches use far more tokens. They're worth it for a market report, not for a quick fact.
16. Be careful with lots of agents at once. When a tool runs several AI agents in parallel, each one reads and writes on its own, so usage multiplies. Use it for genuinely big jobs.
17. Turn repeat jobs into scripts. If the AI does the same check, sync or report every day, it pays for the reading every day. Have it build a script or automation once, and keep the AI for the thinking.
18. Check where your usage goes. Look at your plan's usage page, and at the end of a big job ask the model which steps used the most. You can't fix what you can't see.

A quick checklist before you hit send

• Is this the right project, and a fresh chat for this task?
• Have I said what "done" looks like, in one message?
• Have I pointed to the exact file or page, not "everything"?
• Have I said how long the answer should be?
• Is this a job I keep repeating that should be a skill or a script?

How Promptwise AI sets this up for teams

Most teams don't have a token problem, they have a workflow problem. When we set up AI for a sales, marketing or operations team, we build the structure that makes lean usage the default: projects per department with tight instructions, skills for every repeat process, scripts for the daily checks, and a short playbook so the whole team works the same way. The result is fewer limits hit, lower bills and better output.

Not sure where your team's usage is going? That's exactly the right time to talk it through. Book a free 30-minute chat.

Sources

• OpenAI, What are tokens and how to count them (rules of thumb for English text)
• Anthropic, Claude pricing (input and output prices per million tokens)