Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

Claude Code Caveman Eval: Two Sentences Took the Median Cut From 14.6% to 50.6%

I recounted the repo’s own eval snapshots: the jump came from two added sentences, v3.1.0 still ships the core one, and the same evals swap out Claude Code’s own system prompt.

After the lead, the author explains how Caveman’s eval snapshots were generated and how the “median cut” was reproduced with a script: ten fixed dev questions were run through Claude Code with two reminder variants (a 60-token version and a 106-token version) and a control (“Answer concisely”). The key result is that extending the reminder with two scope-and-length sentences increases the median answer-length reduction from 14.6% to 50.6%, while keeping other factors constant. The article also clarifies that cross-file comparisons (2.9% vs 50.6%) aren’t apples-to-apples due to different skills and prompt setups, that neither eval retains Claude Code’s default system prompt (which materially changes what’s being measured), what exactly changed in the v3.1.0 reminder that ships in the plugin, and how to wire the rule in using the Caveman plugin or by appending a similar scope instruction. Finally, it estimates whether the extra tokens are “worth it” using token/cost arithmetic and ends with recommendations to verify skill version, reminder text, and prompt placement before adopting the reported gains.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI