I came across a Claude Code skill called caveman. The pitch is genuinely funny: make the AI talk like a caveman to save tokens. No articles, no politeness, no “I’d be happy to help with that.” Just fragments. “Bug in middleware. Token check use < not <=. Fix:”

Three intensity levels, a claim of 75% fewer output tokens, and a one-line install — plus a companion tool that compresses your memory files the same way. I wanted to install it.

Token costs add up. Output tokens cost five times more than input tokens, and I have a lot of skills, a lot of memories, a lot of things the assistant has to reason through every time I open a session. A 75% cut on the output side isn’t nothing. The repo was charming, the README was charming, and I was already typing the install command in my head.

Then I ran it through Borg, my assimilation agent. Borg’s whole job is to sit between me and the impulse to install something shiny — it pulls down the skill, looks at what it actually does, and tells me whether it fits my setup before I touch the install command. It exists for exactly this kind of moment.

What came back is the part I want to write down.

It pulled up the repo. It looked at what the skill actually does — inject a system prompt telling the model to be terse. A pattern instruction, nothing more. Then it walked through what would happen if I dropped that into my setup specifically.

The 75% number was cherry-picked. The before/after examples used a baseline of Claude with all the throat-clearing turned on — “Sure! I’d be happy to help with that.” My setup already kills most of that. Realistic savings for me would be closer to 20-30%, not 75%.

The memory-file compression was the dangerous part. My feedback memories aren’t just rules. Each one has a why line and a how to apply line. The why is what lets the assistant judge edge cases instead of blindly following the rule. Compress those into caveman fragments and you save input tokens but lose the reasoning that stops you from making the same mistake twice.

Voice contamination. I’ve spent real effort on a writing voice. A skill that tells the model to talk like a caveman would fight every other writing rule I have. Either it bleeds into drafts or I’m constantly toggling it on and off.

And the kicker: I could get most of the savings with one line in my system prompt. For internal reasoning and status updates, drop the filler. Full voice for anything humans actually read. No third-party skill. No memory-file mutation, no voice bleed.

I didn’t install it. I added the one line to my existing system prompt and moved on.

This is the part I keep underestimating about second brains. The value isn’t ingestion. A system that absorbs everything is just a hoarder with a search bar. The value is having something that knows your setup well enough to push back when ingestion would be a mistake.

I bring something shiny. The assistant says here’s what this would actually do to your setup, and here’s the version that gives you most of the upside without the cost. That conversation is the second brain doing its job.

Most days the system fades into the background — running skills, filing notes, drafting things, answering questions. But once in a while it stops me from doing something I would have regretted, and that’s the day the whole thing pays for itself.

The caveman repo is clever. Almost installing it was the normal impulse. Having something catch me before I did — that’s what the second brain is actually for.