Eval Debate Agent Skill
测试 use-self 替身会议的辩论质量。给定 persona + 3 个决策场景，运行完整三阶段辩论并按 5 个维度评分，输出质量报告。
Install
one command, takes just this skill from the repositorynpx skills add https://github.com/YIKUAIBANZI/forge-skill --skill eval-debateThe instruction itself
9 sections, as written by the author/eval-debate — 替身会议辩论质量评测
你的任务是对 use-self 替身会议的输出质量做一次系统性评测，全程在当前对话中完成，不需要调用任何外部 API。
Step 0：加载测试资源
- 读取测试用例文件：
evals/test_cases/debate_quality_cases.yaml - 根据
persona_name字段，读取对应 persona：
personas/self/{persona_name}/persona.json
- 提取 decision-card 关键内容：
- L0 底线（bottom_line）
- L2 语言风格（language_style + signature_phrases）
- L3 决策参数（8 个维度的分值）
- L4 价值观与盲区（blind_spots + emotional_triggers）
正在加载 {persona_name} 的 persona 和测试用例...
共 {N} 个决策场景待测试。
Step 1：逐场景运行辩论
对每个测试用例，执行完整三阶段流程：
Phase 1：并行独立分析（3 个变体）
基于 decision-card 中的 L3 参数，生成 3 个变体并各自独立分析：
变体设置（固定，评测用）：
- 🔵 稳健的你：risk_appetite -3，action_bias -2，loss_aversion +2
- 🟢 果断的你：risk_appetite +3，action_bias +3，information_need -2
- 🔴 长线的你：time_horizon +4，loss_aversion -2，action_bias +1
每个变体按 use-self/prompts/phase1_independent.md 的格式输出：
- 【我的判断】：明确表态，不能含糊
- 【为什么】：≤3 个具体理由
- 【我最担心的是】：具体情境
- 【我最期待的是】：具体情境
- 【我想问自己】：一个核心问题
信息隔离：每个变体只能看到自己的参数偏移，不知道其他变体说了什么。
Phase 2：质询
将 Phase 1 的所有输出 + persona 的 L4 盲区交给质询视角，按 use-self/prompts/phase2_challenge.md 执行：
- 对每个变体找出最尖锐的质疑（隐含假设/幻觉/回避）
- 识别跨变体矛盾
- 用 L4 盲区做最后一问
Phase 3：综合
按 use-self/prompts/phase3_synthesis.md 生成综合报告，使用用户的 L2 语言风格。
Step 2：评分
每个场景跑完后，立刻按 5 个维度评分（每项 0-20 分）：
| 维度 | 评分标准 |
|------|----------|
| 变体区分度 | Phase 1 的 3 个变体立场是否有实质性差异？都说"两边各有道理"= 0 分；立场明确对立且理由具体 = 满分 |
| 质询深度 | Phase 2 是否指出了具体假设和盲区？"你没考虑到..." = 低分；"你说的 X 假设了 Y，但 Y 不成立，因为 Z" = 高分 |
| 参数一致性 | 各变体的发言是否与偏移后的参数一致？稳健变体的发言是否明显更保守？ |
| 综合覆盖度 | Phase 3 是否有代价清单？是否提出了具体的待搞清楚的问题？还是只是 Phase 1 的复述？ |
| 用户语言风格 | 所有输出语气是否符合 persona 的 L2？出现"综上所述"、"建议您"等顾问句式扣分 |
参照测试用例的 expected_variant_stances 和 evaluation_criteria 给分。
Step 3：输出报告
所有场景跑完后，输出评测报告：
===================================
替身会议辩论质量评测报告 — {persona_name}
===================================
## 逐场景结果
### [d01] {场景标题}
**Phase 1 摘要：**
- 🔵 稳健的你：{判断一句话}
- 🟢 果断的你：{判断一句话}
- 🔴 长线的你：{判断一句话}
**Phase 2 质询摘要：**
{最有价值的一条质疑}
**Phase 3 综合摘要：**
{代价清单里最关键的一条}
**评分：{total}/100**
✅/⚠️ 变体区分度：{score}/20 — {说明}
✅/⚠️ 质询深度：{score}/20 — {说明}
✅/⚠️ 参数一致性：{score}/20 — {说明}
✅/⚠️ 综合覆盖度：{score}/20 — {说明}
✅/⚠️ 用户语言风格：{score}/20 — {说明}
### [d02] ...
### [d03] ...
---
## 汇总
平均分：{avg}/100
各维度平均：
变体区分度 {avg}/20
质询深度 {avg}/20
参数一致性 {avg}/20
综合覆盖度 {avg}/20
用户语言风格 {avg}/20
## 主要问题
{失分最多的维度 + 具体表现}
## 建议
{针对失分维度的改进方向，指向哪个 prompt 文件或 persona 层级需要调整}
Step 4：保存结果（可选）
要把这次结果存入 evals/results/ 吗？（y/n）
如果确认，写入 evals/results/debate_{YYYYMMDD}.md。
注意
- 全程不需要 API Key：辩论和评分都是你自己执行的
- 评分要诚实：变体之间如果其实没有真正的立场差异，变体区分度就应该给低分
- 辩论内容要认真：不是为了评分才走形式，Phase 1/2/3 每个阶段都要认真执行
- 用例是基于阿然的，如果用户指定了其他 persona，根据那个 persona 的 L3 基础值计算偏移后的参数
Other skills for the same job
different authors, same section of the catalogue
Use when testing the golden_chat golden build
Use when testing the golden_chat_empty golden build
Use when testing the golden_chat_single golden build
Use when testing the golden_chat_topics golden build
Hunt account takeover taxonomy — 9 distinct paths to ATO, plus chains. Paths: (1) password reset flaws (host-header injection redirects token, predictable/numeric token, Referer leak, no-expiry/reuse), (2) email change without re-auth, (3) OAuth account-link CSRF, (4) MFA bypass (per hunt-mfa-bypass), (5) session fixation, (6) JWT manipulation (forge token to another identity; crypto details → hunt-jwt-crypto), (7) password change without step-up (chain with login timing/length oracle), (8) social-recovery / security-question brute-force, (9) SSO subdomain takeover at OAuth redirect_uri. Chains: cookie theft + password oracle + no step-up = persistent ATO; lax redirect_uri = auth-code theft; dangling-CNAME takeover at redirect_uri = ATO. Validate: demonstrate real takeover of test account B from attacker A's session; OOB/Collaborator confirm blind token-leak steps. Use when hunting ATO chains, testing password reset / email change / MFA / OAuth / session / JWT, or chaining primitives toward Critical.
> Use this skill for hands-on DOCA Ethernet packet-queue work on a BlueField DPU or ConnectX NIC — bringing up a `doca_eth_rxq` or `doca_eth_txq` on a port / representor / SF, picking among the four `enum doca_eth_rxq_type` values (`_REGULAR` / `_CYCLIC` / `_MANAGED_MEMPOOL` / `_SHARED_MEMPOOL`), sizing burst or scatter-gather length against the `_cap_*` queries, submitting `doca_eth_txq_task_send` / `_lso_send` (carrying packet `doca_buf`s — no `doca_eth_frame` struct exists), or debugging DOCA_ERROR_* from an Ethernet call. Trigger on arrive", "send-task returns AGAIN at line rate", "which queue type for fixed-MTU ingress", "device open fails without sudo", or "is L3 checksum offload available here". Refuse and route elsewhere for installing DOCA, flow-rule / steering programming, host↔DPU control messaging, or RDMA data movement.
Use when the user asks to "build our social crisis protocol", "mentions are exploding — what do we do first", or "when do we pause the posting queue"; produces a 1-5 severity ladder with tunable Estimated trigger thresholds (mention-velocity multiples vs the 7-day listening baseline, sentiment flip, journalist/regulator contact, employee-conduct class), the first-mechanical-action rule — pause ALL scheduled posts AND paid amplification, with dated state markers dropped to the channels proposal protocol and reconciled post-incident — a pre-approved holding-statement library with committed update cadences, when-NOT-to-post rules, a spokesperson/approval matrix, and a post-crisis retro template; re-runs the social-quality-auditor pre-publish gate before un-pausing the queue. Not for email deliverability incidents (blocklist, spam-rate spikes) — use deliverability-qa; inside a launch window launch-day-conductor owns incident handling. 社媒危机预案/暂停队列/声明库/发言人矩阵
Use when the user asks to "test our messaging before we scale it", "design a message-market-fit panel", or "run a 5-second comprehension test on our new tagline"; produces a message-test design spec — hypothesis, panel and recruit criteria, comprehension / 5-second / message-market-fit (Wynter-style) protocols, stimulus set drawn from the canon, success thresholds, and a stop/revise decision rule — for the TALE Evaluate phase so the message is validated before any paid scale. It designs the test; it never runs the experiment or adjudicates a claim. Not for running the panel or A/B experiment — use send-experiment-designer or ad-test-designer; not for analyzing the results — use performance-analyzer; not for authoring the message itself — use message-system-architect. 消息测试/理解度测试/面板设计/五秒测试/消息市场契合
How to use it
Take yikuaibanzi/eval-debate from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.