Unasked for, but I'm going to supply token reduction/best practices for AI tasks that I've learned while using a wide series of different AI providers over the course of my day job:

Feel free to ignore this if you don't care

If you aren't already using something like the Caveman Skill - consider starting. It'll nominally reduce token use, which is always a good thing. It won't make or break the bank, but it's definitely a helpful way to think about how you're sending and receiving your claude input/output.

https://github.com/JuliusBrussee/caveman/releases

Also, consider using something like Obsidian as a vault to store Plans - tell Claude to come up with a plan, use an Obsidian MCP with CLI to print a plan that it can refer to. It saves on repeated context loops, and gives you a documentable track for what it did (and what it was supposed to do).

Also, obviously - don't use the best agents all the time. There's a ton of shit that you can get away with the Good Enough for Gubberment Work style of agent, and using a Plan correctly can really help with the "one agent for all tasks" situation.