Compare two skill outputs blindly and grade expectation evidence.
npx skills add https://github.com/mykcs/myk-skills --skill agents-mykcs
Please help me install this Agent Skill. Skill: agents Source: https://github.com/mykcs/myk-skills/tree/main/agents Command: npx skills add https://github.com/mykcs/myk-skills --skill agents-mykcs
This Skill helps you evaluate competing skill outputs objectively by providing purpose-built agents that analyze results after comparisons, compare outputs without bias, and grade executions against expectations.
Use the comparator agent to judge which of two outputs is better for a specific eval task, then run the analyzer to generate improvement suggestions for the losing skill.
High-intent search queries and answers about installing and using this skill.
Blind comparison evaluates two outputs without knowing their source by generating a scoring rubric, grading structure and content, and selecting a winner based on the rubric outcomes.
Rubric scoring for post-hoc analysis explains why a winner beat a loser by unblinding comparison inputs, inspecting skills, and reading execution transcripts to produce actionable improvement guidance.
Grading expectation evidence as PASS or FAIL involves checking execution transcripts and output files for cited evidence, including detecting superficial or unverifiable claims.
Benchmarking evaluation results require producing JSON outputs for downstream consumption, including winner selection logic and evidence-cited PASS or FAIL grading with optional expectation checks.
Blind benchmarking of multi-run skill outputs prevents bias by generating a rubric, scoring both outputs anonymously, and selecting a winner or declaring a tie based strictly on rubric outcomes.
Execution transcripts and output files are required for evidence-based grading, as the grader checks these sources to verify claims and cite evidence for PASS or FAIL expectation outcomes.
nikhilsitaramBenchmark Claude Code skill revisions with isolated runs and statistical grading.
tnunamakConvert user intent into a trigger-optimized SKILL.md with evaluation workflows.
zonta99Validate ResearchPacks and Implementation Plans against quality rubrics to enforce gating decisions.
enigmaicon-engScore or compare LLM outputs with rubric-based LLM-as-judge evaluation and structured JSON results.
