Welcome to caveman Discussions!

[JuliusBrussee]

🪨 Caveman campfire: benchmarks, feature ideas, reasoning debate & general feedback

---

Caveman just hit the [Hacker News front page](https://news.ycombinator.com/item?id=47647455) and the response has been wild — thank you to everyone who starred, installed, and debated.

Rather than scatter conversations across issues, I want one place to talk about everything. Drop your thoughts under whichever heading fits, or start a new thread.

---

📊 Benchmarks & token savings

The README claims ~75% output token reduction. Some of you (rightfully) want receipts. I'm working on a proper before/after comparison across real Claude Code tasks — file edits, debugging, architecture questions — and will post results here.

If you've measured your own token usage with vs. without caveman, share your numbers. Informal data is still data.

---

🧠 Does brevity hurt reasoning?

This was the biggest debate on HN. The short answer: caveman only affects output tokens, not thinking tokens. Claude still reasons fully before compressing the visible response.

There's also a recent paper ([arXiv:2604.00025](https://arxiv.org/abs/2604.00025)) that found brevity constraints actually improved accuracy by 26 percentage points on certain benchmarks. That said — if you've noticed quality degradation on specific task types, I genuinely want to hear about it. That's the kind of signal that makes the skill better.

---

💡 Feature requests & new personas

A few ideas that came up on HN and Threads:

• Adjustable verbosity levels (full caveman → slightly terse → normal)
• Alternative personas (pirate, haiku, Hemingway, Abathur-from-StarCraft)
• Multilingual caveman — does the prompt work in non-English contexts?
• Auto-toggle — caveman for simple tasks, normal for complex ones

What else would you want? Open to anything — the SKILL.md is just a prompt, so experimenting is cheap.

---

🔥 General feedback

Anything else: bugs, installation issues, hot takes, improvements to the prompt wording, workflows where caveman works surprisingly well (or surprisingly badly). All welcome.

Thanks for making this blow up. Let's make it actually good.

---

[ninsent]

Rocky watch crew die, could not fix. Grace say, Grace will die, rocky fix

[dev-encognitio]

This is actually quite cool! Im wondering if now we can further compress the input by routing every message through caveman compress? On an agentic multi-agent routing system, not interacting with Claude directly and instead through a minifier/summarizer type agent who then also cleans up the context per prompt.

[JustApeasantCoder]

OMG, not only it saves token, it actually makes summaries so much easier to read lol!

[0x0zgdo]

will try this. thanks!

[wzywg]

has anybody eval'd employing caveman as a skill vs an output style?

seems like output style would make alot of sense since it replaces the output style in the system prompt, and be able to pick your levels and turn and on/off with /output-style on the fly.

i'm messing around with output styles, but would love some to see some evals from other people to get some benchmarks.

[TitasDas]

Is there a way I can mimic this by just prompt design and adding a line about how I want the response to be concise ?

[qchrist2]

This repo already optimizes output compression and even instruction-file compression, but I don’t think it yet optimizes for prompt-prefix cacheability. It may be worth restructuring the runtime caveman prompt into a tiny stable core plus a small mode toggle.
Right now risk is: long skill text get re-sent every turn. That overhead can eat token savings.
Better pattern: make caveman small, stable style contract; move changing parts into tiny toggle.
Suggested direction:

• Keep tiny canonical core instruction block.
• Strip runtime skill of examples and explanatory prose.
• Pass only small variable like Mode: lite|full|ultra.
• Keep core wording byte-for-byte stable across turns.
• Put volatile context after stable caveman block, not before.

[cneely33]

Thank you. Should be called Laconic. Best.