Huge difference in token consumption after using Caveman in Claude Code

I have been using Claude Code for the past month. The experience has been excellent with that. I have been able to get a lot of things done with the help of AI automation. It's been a few months since I wrote some real code. Mostly, I get things done with the help of AI. My duty is to review the code and make sure it is correct and intact. Even in my workplace, we have had a huge transition where everyone is supposed to use only AI for coding. Manual coding efforts are not encouraged. This is a big revolution as we are able to do many things within a short span of time.

I have been using Claude Code in my VS Code application. Initially, I used it heavily, and my tokens were exhausted very fast. I was wondering if the issue was on my end or if they were giving us very limited tokens. Later, when looking for some solutions, I came across Caveman through a friend. He suggested using this as a plugin to the extension. I then made a slight research and found out how to use Caveman in my VS Code extension. The plugin claimed to reduce 75 percent of the token consumption. I have to admit that after introducing the plugin, my token consumption drastically reduced.

Source - Gemini

Apart from this, I also learnt some more ways to reduce the token consumption. Another such way is to have a CLAUDE.md file in the repository. So whenever we ask a question to Claude, it gets the basic idea from the CLAUDE.md file and then works accordingly. So some basic questions are automatically sorted. This also reduces a lot of tokens. I asked Claude itself what I should have in the CLAUDE.md file, and it gave me the content by observing the repository.

The next interesting thing is to have a .claudeignore file. Usually, we have .ignore files for many things. The most famous one is for .gitignore, which ignores files for commit. Similar to that, if we have .claudeignore and have appropriate folders and files listed there, Claude ignores them while looking for something. This also saves unnecessary tokens. If we specifically want Claude to check any of the ignored files, we can still mention the file and ask Claude to check it. Overall, the experience has been great, and people who are coding heavily and are desperately looking for a solution to the tokens problem can try the above methods, so that you don't have to pay anything more than your subscription, and it is mostly sufficient for our needs.

If like what I'm doing on Hive, you can vote me as a witness with the links below.