Skip to main content
Most skills teach Claude a task: format this memo, fill this form. This lesson dissects a skill that teaches Claude a behavior instead, and it happens to be one of the funniest files in the ecosystem. The caveman skill is a real community skill with one job: make Claude answer like a smart caveman. Drop the articles, drop the filler, drop the pleasantries, keep every technical fact. Its own frontmatter claims roughly 75 percent token savings, and its body opens with a sentence that is also a demonstration:
Silly on the surface. Underneath, it is a compact masterclass in packaging a behavior, which is why it earns a whole lesson.

​
Feel the compression

The skill defines intensity levels, from professional-but-tight to full classical Chinese. The press below runs the same two answers through every level, with live token counts. The examples at each level are taken from the skill file itself.
Notice what never changes as the dial turns: the diagnosis and the fix. useMemo survives every level. What dies is “Great question!”, “the most likely culprit is”, and eventually the grammar itself. The skill’s whole thesis in one line from its rules section: fluff dies, substance stays.

​
The three moves that make it a good skill

Read past the jokes and the caveman file makes three structural moves that any behavior skill needs. Steal all three. A persistence rule. Behavior skills have a failure mode task skills do not: drift. Claude compresses for three replies, then politeness creeps back. So the skill legislates against it, in caveman:
An escape hatch. Compression is the goal, but not at any price. The skill names the exact situations where clarity outranks brevity, and orders itself out of the way:
A warning about dropping a database table comes out in full, careful English. Then caveman resumes. A behavior skill without an escape hatch is a liability waiting for its edge case. A boundary. The compression applies to chat responses only. Code, commit messages, and pull requests are written normally, because those artifacts outlive the conversation and other humans read them. The skill says so in five words: “Code/commits/PRs: write normal.” Persistence, escape hatch, boundary. That trio is the anatomy of every good behavior skill, whether the behavior is compression, a formal tone for legal drafts, or always-cite-sources mode.

​
Does the money actually matter?

Run the numbers from the press. A verbose answer at roughly 95 tokens compresses to roughly 20 at full caveman. One answer, who cares. But token spend is a rate, not an event:
  • A team member having 40 exchanges a day saves thousands of output tokens daily.
  • In agent workflows where Claude’s outputs feed back in as inputs, shorter outputs compound: every future turn re-reads a smaller history.
  • On metered API work, output tokens are the expensive ones. Cutting them 60 to 75 percent is a real line item.
The deeper habit this builds is noticing that verbosity is a default, not a requirement. You are allowed to change defaults. A skill is the mechanism for changing them permanently instead of pleading “be brief” once per conversation and watching it wear off.

​
Make your own variant in ten minutes

Caveman is a template for any “respond differently, always” skill. To build your own:
  1. Name the behavior in one sentence. “Answers include a confidence level and the single strongest counterargument.”
  2. Write five rules that produce it, with a before and after example pair. Examples teach Claude the register faster than descriptions do.
  3. Add the trio: a persistence rule, an escape hatch for the cases where the behavior would hurt, and a boundary naming where it does not apply.
  4. Put the trigger phrases in the description: the exact words you would naturally say when you want this mode.
The lab at the end of this section walks the full install-and-test loop if you want to ship one today.

The caveman skill is an open community skill; the excerpts and level examples above are quoted from its SKILL.md as installed. Token counts in the press are approximate, at the standard rough rate of one token per three quarters of a word.