AgentSkillsCN

win-evaluate

邀请 WIN 委员会专家对内容变体进行评分。可用于生成 committee_matrix.json，比较各草案，并调整评分权重或专家关注重点。

SKILL.md
--- frontmatter
name: win-evaluate
description: Score content variants with WIN committee experts. Use to generate committee_matrix.json, compare drafts, and tune rubric weights or expert focus.

WIN Evaluate

Overview

Run only the committee scoring stage to produce data/output/committee_matrix.json.

Workflow

  1. •Ensure profiles.json already exists (run profile stage first if needed).
  2. •Validate content and committee input contracts.
  3. •If required content/committee inputs are missing, please interview the user to acquire the necessary inputs before running commands.
  4. •Run committee evaluation and inspect weighted scores plus consensus spread.

Command

bash
scripts/run_evaluate.sh

Input Contracts

Expected schemas are documented in:

  • •../win-committee/references/data-contracts.md

Guardrails

  • •Keep scoring criteria explicit through rubric_weights.
  • •Avoid overfitting to one persona by checking expert weights and consensus spread.
  • •If required inputs are missing, please interview the user to acquire the necessary inputs.
  • •Keep user-specific drafts in *.local.json files.