I burned a lot of tokens using Claude Code until I learned how to use plan mode. Not entirely groundbreaking news here, but I didn't start out the right way when I first dived into Claude Code and it cost me hundreds of dollars in tokens.
Prior to Claude Code, I was using Claude API via Hermes. I got into the habit of burning tokens because I didn't see any other way to do it. Every chat, command and action I asked Hermes to do, it did so eagerly while burning tokens via Claude API.
Colleagues kept telling me about Claude Code, which has been around for a while. But I was stubborn. Why stop now when everything I needed to do, I was already able to do using my Hermes agentic AI?
And then the frustration finally reached a tipping point and I got tired of buying more credits. There had to be a better way. So I took a step backwards, installed Claude Code, and started using it on my terminal.
I discovered that it did everything my Hermes and Claude setup did, and with the pro subscription I wasn't paying anything more than the subscription fee... until Claude Code told me I was at the usage limit and I had to buy more credits. Which I did, eagerly, having discovered a new way to build my websites and apps. I gave it $75 Canadian... and I quickly burned through that as well as I voraciously typed and built. And this was just day one.
Then I did what every token-burning-addicted Claude Code user did... and bought more tokens. Except I had enough and I decided to do the obvious thing... which was to learn about why I was paying so much to use this thing. I took Claude Code 101 where I learned about how conversations can be /clear (ed) and /compact (ed). I finally understood what it meant to push the context windows. Beyond that, I finally understood why I was burning so many tokens.
I shifted Claude Code into plan mode before tinkering with code generation, and that helped a lot. AI can easily get overwhelmed and take a tangent without proper planning (which should've been obvious from the beginning). I also kept an eye on my usage levels. I realized that by using plan mode, I didn't even have to /clear or /compact because plan mode efficiently structured the concept of what I was building and it approached it without wasting tokens.
Yes, there were hiccups along the way, and Claude Code isn't perfect. Nor am I an experienced enough coder/developer to guide it properly. But I haven't burned overage credits since learning how to use Claude Code properly and now I'm eagerly looking for more ways to build efficiently and to prevent overage.
Long story short: learn how to use Claude Code before going keyboard happy.