Last week a client texted me a screenshot of her Claude usage meter, hit her limit by 2pm, and said “I asked like ten questions today. How is this possible?”

I went looking for the answer, and it led me to a piece by a builder named Nate who tracks his own AI token usage obsessively. He opened his tracker after a long day and found 3.77 billion tokens. Most of it was “reused” input: material he never typed.

Why? Because every time you message the model, the LLM is reloading all the previous context as well as the new message you just sent.

So the tenth message in a conversation can be far bigger than the first, even though what you typed looks shorter. You’re not just sending your new sentence. You’re sending everything that came before it, every time.

That’s not a reason to panic about your AI bill. It’s a reason to get more deliberate about what you’re dragging along with you. A few habits I’ve started using myself and with clients, inspired by Nate's list and consolidated for those that resonate with knowledge workers:

1. Hit rewind. If AI misreads your request and you correct it, that correction usually gets added on top of the original mistake, forever, in every future message. If the wrong turn doesn’t teach you anything, edit the original message or start over instead of layering a fix on a mistake. If the mistake itself was useful (it revealed a real misunderstanding), keep it. Otherwise, let it go.

2. Batch what belongs together. If you need a summary, the risks, and five headline options from the same document, ask for all three in one message. Three separate questions often mean the AI reloads the whole source and every prior answer, three separate times. One ask, one load.

3. Start clean when the job changes, not when you feel like it. I used to keep threads running for days out of habit. Now the test is simple: is this still the same job? If yes, keep going; the AI needs that history. If the job has actually changed, a fresh thread is faster and cheaper than dragging three unrelated projects behind you.

4. Carry the answer, not the argument. If a document took six drafts to get right, your next request only needs draft six and what you want changed, not the five rejected versions and every round of feedback. I now save the accepted version somewhere I can grab it, and build the next ask from that plus the change. Bonus: this also stops AI from accidentally resurrecting an idea you already killed three drafts ago.

5. Ask for what you actually need, not a performance of thinking. If you need one paragraph, ask for one paragraph. If you need three bullet points for a slide, don’t ask for “some thoughts on this.”

None of these require a new tool. They’re just a different question to ask yourself before you hit send: does the AI actually need this to do the next thing well, or is it just… along for the ride?

Alex

PS if you want the deeper breakdown this came from (measurement methodology and all), it’s all in Nate’s YouTube video. Hit reply and I'll send you the 20 minute video.