Caveman Output Style for Claude Code

Better consistency and adherence than CLAUDE.md or skills. Always-on formatting. Less fluff, same technical signal.

quick start

One command. Done.

# macOS / Linux
mkdir -p ~/.claude/output-styles && cp ./.claude/output-styles/caveman.md ~/.claude/output-styles/caveman.md

# Windows
New-Item -ItemType Directory $HOME\.claude\output-styles -Force | Out-Null; Copy-Item .\.claude\output-styles\caveman.md $HOME\.claude\output-styles\caveman.md -Force

Then: /config → Output Style → Caveman.

Want it always on? Add this to ~/.claude/settings.json:

{ "outputStyle": "caveman" }

Why use caveman?

prompt	normal answer	caveman answer	rough token cut
why build fail?	Build is failing because TypeScript cannot resolve the @/auth import after the alias change. Update the path mapping or import path.	Build fails: TypeScript cannot resolve @/auth. Fix alias or import path.	~42%
what changed in auth?	I updated the middleware to reject missing bearer tokens before decoding the JWT, which prevents the null token crash.	Auth middleware now rejects missing bearer token before JWT decode. Crash gone.	~45%
how set default style?	Open your Claude settings file and add the outputStyle field with the value caveman, then start a new session.	Add "outputStyle": "caveman" to settings. Start new session.	~47%

Real token count will vary. Table shows rough reduction only. Pattern is the same: less fluff out -> less output spend.

Same answer, less bla. Shorter replies every turn, lower token bill and less energy wasted on filler.

Use when:

• You keep typing "be concise"
• You want the same terse format on every turn
• You want less output spend without losing technical substance

Backed by research: Brevity Constraints Reverse Performance Hierarchies in Language Models. Shorter answer keeps quality, and sometimes even improves it.

why output style, not other things

Claude Code official documentation show output styles change how Claude responds, not what Claude knows.

Output styles are the perfect matching tool for achieving "talk short on every turn" goal:

tool	good for	why worse?
CLAUDE.md	project rules, conventions, codebase memory	adds user message after default prompt. not purpose-built for response style, directive lost on long threads
skills	reusable workflows, task prompts	skills load when invoked or relevant. Claude Code may decide against using it. Not always-on formatting layer
--append-system-prompt	one-off system prompt append	appended at end of system prompt. lower priority. busts cache if value changes

All methods add roughly the same amount of input tokens. Output styles win on adherence and consistency, not input cost.

how it works

Output styles inject a dedicated section into Claude Code's system prompt:

# Output Style: Caveman
[style text here]

This section is:

• Dedicated — purpose-built for response formatting
• System prompt level — not buried in user messages like CLAUDE.md
• Always on — no model discretion, no invocation needed
• Survives long threads — style persists without dilution

available styles

Caveman (default)

Standard terse mode. Drop articles, filler, pleasantries, hedging. Fragments OK.

Caveman Ultra

Maximum terseness. Telegraphic fragments. Symbols over words. Abbreviated prose. Use when you want the most aggressive compression.

# macOS / Linux
cp ./.claude/output-styles/caveman-ultra.md ~/.claude/output-styles/caveman-ultra.md

# Windows
Copy-Item .\.claude\output-styles\caveman-ultra.md $HOME\.claude\output-styles\caveman-ultra.md -Force

Then: /config → Output Style → caveman-ultra.

Or set it in settings.json:

{ "outputStyle": "caveman-ultra" }

Tokenizer-safe terseness

Caveman (default) avoids aggressive abbreviations that may hurt clarity or tokenization on non-Anthropic models. Caveman Ultra goes further with symbols and prose abbreviations — use it when you trust your model's tokenizer.

FAQ

Does caveman make Claude dumber?

Not at all. Same knowledge, same reasoning, just shorter output. Research shows output brevity can even improve answer quality.

Will it break code generation?

No. keep-coding-instructions: true keeps all coding behavior intact. Only the voice changes.

Can I switch back?

Sure! /options → pick different style. Or remove it from settings.json.

Does it work with skills and agents?

Yes. Output style is the base voice; skills and agents still work normally.

Why not just tell Claude "be concise"?

You can, but you have to say it every. single. turn. Output style is always on, set once and forget.

Does it save money?

Yes. Fewer output tokens = lower cost, especially on long sessions.

Can I customize it?

You are more than welcome. Edit caveman.md in your ~/.claude/output-styles/ folder and make it your own.

license

MIT. Do what you wish with it.

---

Like this? Star the repo. Share with your team.