INFYNY LABSWhat you can save
If Claude or Codex gives you a long answer when you only need the result, Caveman can help it use fewer words. Fewer words will usually need fewer reply tokens.
- Tokens are small pieces of text. They are not always whole words.
- This post covers the Caveman answer-style skill. The repository also contains a larger suite with other features.
- The skill asks AI to answer first, remove filler and keep important details.
The practical goal is fewer unnecessary reply tokens, with an answer you can still use.
The two-person example
Imagine two colleagues asking for the same internal order summary. The posters show a longer and shorter way to say the same thing.
- The order has two blue chairs. Delivery is Friday. Payment has not been received.
- The short example keeps each of those facts without the extra sentences.
- We wrote both examples. The people and token strips are illustrations, not measured Caveman outputs or a benchmark.
Compare meaning as well as length. Shorter is useful only if the details survive.
Pros and cons
Caveman suits quick internal summaries and direct work answers. It is less useful when you need patient teaching or a careful explanation.
- Pros: less filler, fewer reply tokens when the reply is shorter, and less text to scan.
- Cons: compressed answers can need follow-up questions, and the skill cannot guarantee correctness.
- Ask for more context when you need it. Important numbers, dates, negations and warnings should remain.
Use a short answer for a quick result. Ask for a full explanation for a difficult decision.
How to try it in Claude or Codex
Read the repository instructions before installing. Its skill-only option is separate from installing the whole Caveman suite.
- The README lists: npx skills add JuliusBrussee/caveman -g
- After installation, invoke /caveman in Claude or $caveman in Codex.
- Say stop caveman to return to normal replies.
- Try an ordinary internal note without private customer information.
- Check the shorter answer against the original. Ask for detail if anything is unclear.
We inspected the instructions; we did not install or benchmark the skill for this post.
Reply tokens and your bill
Saving reply tokens and saving money are different things. Total usage includes the text AI reads, the skill instructions and tool activity as well as its replies.
- The amount saved depends on the request, answer and model. We are not promising a percentage.
- The skill itself adds instructions. Extra follow-up questions may also use tokens.
- A fixed-price subscription may not become cheaper because a reply is shorter.
Comment CAVEMAN to receive only https://github.com/juliusbrussee/caveman. Setup and licensing details are available there.