# Caveman vs RTK vs plain Claude Code on the same bug fix. Neither cut total tokens much — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/d21a4076c85c5fd2

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-11-02 11:58 ingested 2026-11-03 02:42

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1xlv8wg/caveman_vs_rtk_vs_plain_claude_code_on_the_same/)

## retained text



```
Both of these are meant to cut token usage, so they were run against plain Claude Code on the same small task. A Python project with 79 tests and one planted bug (`>` instead of `>=`). The prompt: run pytest, find the bug, make the smallest fix, rerun, show git diff. Claude Code 2.1.316, Sonnet 5.5, two runs per setup, fresh copy each time.

* RTK (rtk-ai/rtk) rewrites Bash commands like `pytest` into `rtk pytest` and shortens the output Claude reads. It only works on Bash calls, not Claude’s own Read or Grep tools
* Caveman (JuliusBrussee/caveman, plugin version) adds rules at session start asking Claude for shorter replies

Averages, with costs from Claude Code’s own estimates:

* plain: 165,122 input tokens, about $0.052
* Caveman: 164,852 input tokens, about $0.065
* RTK: 162,774 input tokens, about $0.049 (second run only)

RTK cut tool output by 46%, and verbose pytest output went from 10,291 characters to 974 with the failures still shown. But total input only dropped about 1.4%. Part of the reason: plain Claude was already piping pytest through `tail`, so there wasn’t much left for RTK to save.

RTK’s first run cost $0.155. It ran first, so it had to write the prompt cache from scratch, which is why only its second run is used for cost. That’s one run, so not enough to call it a real saving.

Caveman’s final messages were a bit shorter (71 and 91 words vs 94 and 94), but output tokens barely moved since the code itself doesn’t get shorter. It cost about 25% more here. It also wrote about 3,400 more cache tokens per run, likely from the rules it loads at session start. Its own two runs varied more than the gap between setups, so take that with salt.

All six runs found the bug, made the same two-line fix and ended with 79 passing tests.

One small task, six runs total. Longer sessions or projects with lots of test or log output could look different, and Caveman’s CLI features weren’t tested.

[https://aiparabellum.com/caveman-vs-rtk-claude-code-token-test/](https://aiparabellum.com/caveman-vs-rtk-claude-code-token-test/)
```



document fetched 2026-11-03 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1xlv8wg/caveman_vs_rtk_vs_plain_claude_code_on_the_same
