File language: Not specified
Save Tokens with caveman
Reduce response tokens by an average of 65% while preserving accuracy.
에반Free
Skill overview
Trying to save AI tokens?
One command: caveman.
Automatic in Claude, Codex and Gemini.
You get responses like this every day.
As models get smarter, answers get longer.
A one-sentence answer becomes five paragraphs.
-
01Another “Of course!” followed by 1,200 tokens.
You read five paragraphs for a one-sentence answer. Every response has the same verbose tone.
-
02The context window is already strained halfway through.
Long replies accumulate until a session can't last the whole task. Restarting interrupts the flow.
-
03Constantly checking token usage?
Keeping the usage dashboard open has become routine. You top up before reaching the limit.
The model isn't the missing piece. A tone trained to answer at length..
Fewer tokens.
The information you need stays.
/caveman
One invocation compresses the output, leaving the reasoning process unchanged.
Three ways to reduce tokens.
Average response-token reduction: 65%
The same question to the same AI: a typical answer uses 1,214 tokens. With caveman, it uses 294. The source benchmark reports the same information and accuracy, with filler removed.
Memory files compressed by 46% too.
Rewrite CLAUDE.md, project notes and memory files loaded every session. Average size falls from 898 to 481 tokens, preserving information and behavior while reducing input tokens.
Install once. Use across thirty tools.
Automatic in Claude Code, Codex and Gemini. Installation includes rule files for thirty-plus agents including Cursor, Windsurf, Cline and Copilot, keeping the response style consistent.
Three folders. That's all.
MIT open source. Installation uses this exact structure.
Benchmark reproduction scripts and the evaluation harness are available to inspect.
caveman/ ├── skills/ — /caveman, /caveman-commit, /caveman-review ├── benchmarks/ — Token benchmarks across 10 tasks + reproduction script └── evals/ — 3-arm evaluation harness: baseline · terse · skill
Here's what the answer looks like.
The same question to the same AI. The content is the same; the length differs.
" I'm happy to help! This is a very common problem, and it can certainly be solved. To start with the conclusion, it appears to come from the time comparison in your code. Looking more closely, when comparing the expiration time and current time, < you used the operator , which means behavior may differ from your intent when the times are exactly equal. To resolve this, < So we'll show <= change it as recommended. Please let me know if you have any other questions!"
Filler caveman removes
“Auth middleware bug. Token-expiration check < uses <= Change it to this.”
Frequently asked questions.
“More tokens wasted.”
Start reducing that waste.
Q&A 0
No questions yet.
From purchase to your first result
Check what you need in the guide, then start a first task with your own materials.
Open the guide for your skill
Open the guide from the payment confirmation or your library, and check the accounts and materials you need.
Get the skill and follow its setup guide
Open the skill source link and follow the accompanying setup instructions.
Try a small task with your own materials
Start with one topic or item. Describe the result you want and check whether it works for your needs.
Ask in the skill room if you get stuck
Share the task, the step where you got stuck, and any error message in the buyer community. You can return to the guide from your library.
Example first request
“I’d like to try a first task with Save Tokens with caveman. Tell me what materials and setup I need, then help me start with a small task.”
Purchase information
Before you buy
- Materials
Skill source link
Open purchased materials from your library.
- Requirements
You need an AI agent and an account to run the skill.
Check the product description for supported agents and required services.
- Additional costs
External AI services may have separate subscription or usage fees.
Ask the seller about required plans and expected usage before buying.
Warranty and support
Review the tested environment and support period.
- Tested environment
- Tested with Claude Code and Codex. Operation with other agents or CLIs is not guaranteed.
- Free patch period
- One year from upload (2026-05-11), through 2027-05-11. If model or CLI changes break the skill during this period, we will fix it at no charge.
- Access period
- Available to download immediately after payment, and available to download and view again in your library for 11 months from the payment date. Download and viewing access ends after this period.
- If we cannot fix it
- If we cannot fix it within the free patch period, we refund an amount proportional to the remaining period.
- Last verified
- 2026-10-24(skill's last updated date)