Caveman

Fewer words. Fewer tokens. Bigger savings.

Every AI coding session has a hidden tax. It shows up in your bill at the end of the month and most developers never trace it back to the source.

The culprit isn't the code generation. It's the words around it.

**AI models are trained to be polite.** They open with greetings. They hedge with "I would recommend." They close with summaries of what they just said. In a single session, these pleasantries stack up fast. Multiply that across a full workday and you are burning thousands of tokens on prose that carries zero technical signal.

Julius Brussee built Caveman to fix this. It's an open-source system prompt modifier that strips the verbal fat from AI responses. Articles go (a, an, the). Filler phrases go. Hedging goes. What remains is dense, direct technical information: the actual answer.

The tool ships three intensity levels. Lite removes obvious filler but keeps full sentences. Full drops articles and switches to short fragments. Ultra compresses everything into near-raw notation where a security issue becomes "auth mw throws 401, add refresh path" instead of a three-paragraph explanation.

**The compound effect is where it gets interesting.**

Coding agents don't just answer your current question. They re-read the entire conversation history on every single turn. Every verbose response you received on turn two is still burning tokens on turn nine. Terse replies keep the context window lean, and a lean context window compounds across the session. The savings aren't linear.

Caveman handles this by being surgical. Code blocks, file paths, variable names, and security warnings are preserved byte-for-byte. The accuracy doesn't drop. Only the prose dies.

The installer is one line. The payoff is immediate.