Written in caveman mode. Translation below.
Me experiment with caveman skill for Claude Code. Skill cut filler. Drop article, drop hedge, drop pleasantry. Technical substance stay. Token count drop.
Normal Claude response: "Sure! I'd be happy to help you implement a solution for that. The issue you're experiencing is likely caused by..." — many word, little info.
Caveman response: "Bug in auth middleware. Token expiry check use < not <=. Fix:" — few word, same info.
Session get long → context fill → cost rise. Caveman fight this. Same meaning, fewer token. Signal survive, noise die.
Me try on real work. Code review, bug fix, planning. Work well. No lose precision. Just lose fluff.
Efficient tool for efficient human. Me keep using.
Translation
I spend a lot of time in long Claude Code sessions. Every message costs tokens and long sessions fill context windows fast. At some point you start paying for filler.
I found a skill called Caveman that changes how I write to the AI. It drops articles, filler and pleasantries. It keeps the technical meaning. A sentence like "Sure, I'd be happy to help you implement a solution for that" becomes "Bug. Fix:". The result is the same. The padding is gone.
Caveman cuts roughly 60 to 75 percent of token count without losing meaning. I tested it on code reviews, debugging sessions and planning runs. No precision lost.
As a developer trying to get more done with less friction, it fits. I will keep using it.