Your CLAUDE.md Is 46% Filler. Caveman Proves It.

I ran the caveman skill on my project's CLAUDE.md last week expecting to lose something.

I didn't. The agent performed the same. In some cases, better — fewer instructions seemed to reduce the chance of contradictions the model had to silently arbitrate.

That result disturbed me more than a performance drop would have.

What Caveman Actually Does

The caveman skill from juliusbrussee/caveman does one thing: it rewrites your prompts and instruction files in stripped-down, primitive English. No caveats. No politeness. No "please make sure to". No "it's important that". Just: "Do X. Not Y. If Z, then W."

It's named after the language register it targets — blunt, noun-heavy, verb-forward. The kind of instructions a caveman might give if a caveman were an extremely competent senior engineer.

The skill documents that it typically achieves 40–60% token reduction without loss of agent behaviour. That's not a rounding error. That's almost half of what you're sending, every single context window, for every task.

The Math on Wasted Tokens

Context windows have gotten large enough that most people have stopped thinking about what goes in them. But large context windows don't change the cost structure — they change the threshold at which the cost becomes visible.

At 92,000 installs, caveman is one of the most-installed skills in the ecosystem. That number suggests a lot of people have already run the experiment and confirmed the result: their instructions were bloated.

The tokens you're burning on hedging language and polite framing are not improving comprehension. They're noise. The model has to process them anyway.

More insidiously, verbose instructions tend to bury important constraints in padding. The model doesn't experience your CLAUDE.md the way you read it — front-to-back, with emphasis where you put it. It processes all of it at once, weighted by what follows. When you say "please be careful to make sure that you avoid" before the thing you actually don't want, the thing you don't want has less weight than you think.

What You Find Out When You Compress

The uncomfortable discovery — the one that made me sit with the result for longer than I expected — is this: compressing a CLAUDE.md by 46% and seeing identical performance means the removed 46% wasn't doing anything.

But I'd written it. I'd thought about it. I'd put it there on purpose.

What happened? A few things, on reflection:

Some instructions were written to reassure myself rather than to guide the agent. "Remember to check for TypeScript errors before committing" is the kind of instruction you write after a bad experience. It made me feel like the problem was handled. It didn't add information the agent didn't already have.

Some instructions were outdated — rules written for an earlier version of the project that no longer applied, left in because nobody had reviewed the file holistically.

Some were redundant — three different sections that all said variations of the same thing because they'd been written at different times by someone (me) who didn't remember what they'd already written.

Caveman found all of this. Not by auditing it — by forcing everything into a format where redundancy and noise become obvious.

The Discipline Is the Point

What I've started doing is running caveman on every new instruction file I write, before I commit it. Not to minimize tokens as a goal in itself — but because the compression process surfaces unclear thinking.

If an instruction resists caveman compression — if making it primitive makes it lose meaning — that usually means it was trying to do two things at once, or it was a vague directive masquerading as a specific rule.

The constraint of "say it like a caveman" is, paradoxically, a precision tool. You can't be vague in minimal English. You have to know what you actually want.

That discipline belongs in the prompt writing process from the start, not as a cleanup pass at the end. If you can't say it in ten words, you probably haven't decided what you're asking for.

---

Part of the AI Skill Daily series — skills worth understanding, one at a time.