ultracave skill: make Claude answer in the shortest fragments

Makes Claude answer in bare fragments with each fact said once, until you say "stop caveman".

Installs: 33K; on skills.sh
Stars: 116.7K; on GitHub
Files: 1; besides SKILL.md
Updated: Nov 2026; last repo commit

ultracave is the maximum-compression mode of the caveman skills. Claude drops articles and connecting words, uses one word where one will do, and states each fact once. Code, commands, paths, error text, numbers and words like not and never are never cut.

It starts only when you type /ultracave or /caveman ultra, and stays on until you say "stop caveman" or "normal mode". It switches back to full sentences for security warnings, irreversible actions and anything saved outside the chat, such as commits and docs.

Good for

Developers who read fast and want the answer with no padding
People who pay per token and want shorter replies
Anyone tired of long, polite preambles from an assistant

Install ultracave

Claude Code

Copy this prompt.

Install the ultracave skill in Claude Code. Run this command:

npx skills add https://github.com/juliusbrussee/caveman --skill ultracave

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

From its listing on skills.sh. Source.

What's in the skill

When it loads

Caveman at maximum compression: fragments, one word when one word is enough, each fact once. Invoke only with /ultracave or /caveman ultra. Stays on until "stop caveman" or "normal mode".

Sections in SKILL.md

Persistence
Floor
Rules
1. Fragments
2. One word when one word is enough
3. Each fact once
4. Cut conjunctions only when order survives
5. Tool runs
6. Never perform
When to break the rules
Pre-send check

Key facts

Author: juliusbrussee
Repo: juliusbrussee/caveman
Installs on skills.sh: 33,008 as of Nov 2, 2026
GitHub stars: 116,702 for the whole juliusbrussee/caveman repo, Nov 2, 2026
License: Apache-2.0 (crawl data; the repo README also says Apache-2.0, whole repo)
Last repo commit: Nov 1, 2026
What is inside: 2 files: SKILL.md and README.md
skills.sh audits: Gen Agent Trust Hub pass, Socket pass, Snyk pass
How it starts: Only when you type /ultracave (or its /caveman alias); the SKILL.md says to invoke it only that way
Stays on: Until you say "stop caveman" or "normal mode"
Part of: The caveman repo, next to the base caveman skill

What it includes

Turns on the fragment voice for the whole session
Also reachable as /caveman ultra, which the README calls an alias.
Try: “/ultracave”
/ultracave
Reports the current mode and changes nothing
Without a hook it reports the mode you were in, or off, plus a note that the host does not track it.
/caveman status
Drop articles and connecting words when the order stays clear
Rule 1. Example: "Inline object prop, new ref, re-render. useMemo."
Fragments
"Yes. Null check first." instead of a hedged sentence
Rule 2.
One word when one word is enough
No restating, no summary after a list
Rule 3. For code changes it shows changed lines plus context, not the whole file.
Each fact once
Keeps steps in order even when words are cut
Rule 4.
Cut conjunctions only when order survives
Says nothing between tool calls unless something changes
Rule 5. Exceptions are a change of direction, a needed confirmation, or a security or irreversible step ahead.
Tool runs: one line in, one line out
Code, commands, paths, API names, errors, numbers and negations stay verbatim
Also kept: the user's language, one term per thing, and existing code comments.
The floor
Security warnings, irreversible actions, unclear fragments, questions to you
Also anything persisted outside chat: code, comments, commits, docs, issues, PRs, tickets, memory and third-party messages.
When it speaks plainly
Say "stop caveman" or "normal mode" to end it
Stop

From the official docs

Questions

What does ultracave do?

It makes Claude reply in bare fragments with each fact said once. Code, commands, paths, error messages, numbers and negations such as not and never stay exact.

How do I install ultracave?

Use the install box at the top of this page. Then type /ultracave to start. The /caveman ultra shortcut comes from the base caveman skill, so it works only if that is installed too.

How do I turn ultracave off?

Say "stop caveman" or "normal mode". The skill stays on for every reply until you do, and it stays on when Claude is unsure.

ultracave vs caveman: which should I use?

The skill's README calls ultracave "Caveman with the grammar stripped." Use caveman for short but readable answers. Use ultracave when you read fast and want the payload only. The README says to skip ultracave for onboarding, security review and anything other people will read.

Does ultracave really save tokens?

The README reports 35% fewer output tokens at the median than a plain "Answer concisely." instruction. That was measured once on ten prompts with one model and one run, and it counts length only, not correctness. The main README adds that the rules cost about 1,160 tokens of input, and that shorter answers save nothing if you pay per request, as with GitHub Copilot.

Does ultracave work outside Claude Code?

Its README documents /ultracave in Claude Code. The main caveman README says its installer also sets up Codex, Gemini CLI, OpenCode and more, and that Codex uses $caveman. It does not say $ultracave.

Is ultracave safe?

The skill is two text files, SKILL.md and README.md, and bundles no scripts. The main caveman README says the skill runs on your machine and sends nothing, and that its separate command line installer sends usage stats unless you turn them off with caveman telemetry off or DO_NOT_TRACK=1.

Information

updated Nov 2, 2026

Repository: juliusbrussee/caveman
Folder: skills/ultracave

juliusbrussee/caveman skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 3

ultracave SKILL.md · github.com
ultracave README · github.com
skills.sh listing · skills.sh