evallisted
Install: claude install-skill x-mesh/xm
<Purpose>
x-eval structurally evaluates the quality of AI agent outputs. Independent judge agents score outputs using multi-dimensional rubrics, supporting strategy/model benchmarking and A/B prompt experiments.
No external dependencies. Evaluation results are stored in `.xm/eval/`.
</Purpose>
<Use_When>
- User wants to score or grade agent output against a rubric
- User says "evaluate", "grade", "quality check", "score", "eval"
- User wants to compare two outputs and pick the better one
- User says "compare", "which is better?", "A vs B"
- User wants to benchmark strategies or models on the same task
- User says "bench", "benchmark", "which strategy is better?"
- User wants to create or list evaluation rubrics
- User says "rubric", "evaluation criteria"
</Use_When>
<Do_Not_Use_When>
- Simple factual questions that don't need quality evaluation
- Strategy execution without evaluation (use x-op instead)
- Project lifecycle management (use x-build instead)
</Do_Not_Use_When>
# x-eval — Agent Output Quality Evaluation
Multi-dimensional rubric scoring, strategy benchmarking, A/B prompt experiments.
Judge agents fan out for independent evaluation, then aggregate results.
## Mode Detection
Read mode from `.xm/config.json` (`mode` field). Default: `developer`.
**Developer mode**: Use technical terms (rubric, benchmark, verdict, score, dimension). Concise.
**Normal mode**: Guide in plain, accessible language.
- "rubric" → "평가 기준", "benchmark" → "비교 평가", "verdict" → "판정", "dim