360 Secure SkillHub
360 安全技能中心

Debate Learning Workflow

规定每个主题至少进行 20 轮、最多 40 轮的证据型辩论,开场须包含主张、证据与来源,每轮记录发现的漏洞、成因与下一轮修复方案,由裁判把关是否继续;输出主题记录、每日索引和经验追加文件,并对每条通用化经验做七项完整性校验。

kdylan1010-alt
kdylan1010-alt
热度 46.93
最近更新 2026-05-11 17:35:58
分类:知识依赖:内部调用大模型 / 读写文件应用场景:研究 / 笔记整理 / 概念解释 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add debate-learning-workflow
Metadata
name
debate-learning-workflow
description
Run evidence-backed multi-agent debates (A/B/Opponent3/Judge) with 20-40 rounds, loophole analysis, and universal actionable lesson extraction.

Debate Learning Workflow

Use this skill to run structured, high-rigor debates that produce transferable learning.

Core Rules

  • Minimum 20 rounds per topic.
  • Judge-gated continuation if critical loopholes remain.
  • Hard stop at 40 rounds.
  • Openings must include Claim + Evidence + Source.
  • Unsupported openings are invalid and must be revised.
  • Every round logs:
    • loophole found
    • why loophole exists
    • concrete fix for next round

Roles

  • Debater A
  • Debater B
  • Opponent3 (alternative/challenger model)
  • Judge (evidence quality + unresolved uncertainty)

Output Files

  • Topic file: ~/Desktop/debate/YYYY-MM-DD-topic.md
  • Daily index: ~/Desktop/debate/YYYY-MM-DD-index.md
  • Lessons append target: ~/Desktop/lessons.md

Universalization Quality Gate

Every generalized lesson must include:

  1. Trigger condition
  2. Loophole/failure pattern
  3. Root cause
  4. Corrective action
  5. Measurable metric/threshold
  6. Boundary conditions
  7. Transfer examples in at least 2 other fields

If any field is missing, lesson is INVALID and must be rewritten.