Developers often tweak prompts hoping for better results, but it's hard to know if changes truly work or just produce different outputs due to LLM variance. The new rashomon plugin for Claude Code tackles this by running both original and optimized prompts in isolated environments, providing concrete evidence of improvement rather than subjective guesswork.

For instance, a vague prompt like "Add logging to track function calling usage" was analyzed and optimized with specific context, requirements, and output formats. The tool flagged missing security considerations and unclear scopes, transforming a superficial request into a structural improvement that covered both detection and execution phases for a Slack bot using Google Gemini API.

This approach moves prompt engineering from trial-and-error to a measurable process. rashomon distinguishes between meaningful changes and mere variance, helping teams iterate with confidence. It's designed for daily Claude Code users who need evidence, not vibes, when refining complex prompts for coding, analysis, or writing tasks.