Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add YIKUAIBANZI/forge-skill --skill eval-debategit clone --depth 1 https://github.com/YIKUAIBANZI/forge-skillWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/yikuaibanzi/forge-skill/eval-debate)<a href="https://agentmods.dev/skills/yikuaibanzi/forge-skill/eval-debate"><img src="https://agentmods.dev/badge/skills/yikuaibanzi/forge-skill/eval-debate/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/yikuaibanzi/forge-skill/eval-debate"><img src="https://agentmods.dev/badge/skills/yikuaibanzi/forge-skill/eval-debate.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00049 | $0.01564 |
| Opus 5 | $0.00024 | $0.00782 |
| Sonnet 5 | $0.00010 | $0.00313 |
| Haiku 4.5 | $0.00005 | $0.00156 |
Grade A, and why
eval-debate scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 158 lines — stays where its author put it; the contents beside it link to each section on GitHub.
/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}
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 9d ago First seen · 158 lines · 49 tokens per session scan A ae5412f55227
eval-debate is a skill published in the GitHub repository YIKUAIBANZI/forge-skill (116 stars, last pushed 5mo ago), licensed MIT. It adds 49 tokens to every session and 1,564 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
weak-agent-test
Run the weak-agent adversarial test harness against docx-cli. Spawns weak exercise agents (Haiku by default, Sonnet to probe, or a local agent harness's pre-produced runs) to perform real document tasks over six scenarios — five editing (MNDA form-fill + font fidelity, invoice table-edit/restructure + logo replace…
declaring-invariants
Find tests that enumerate a domain by copying it, and declare the invariants a codebase depends on. Reports where a parametrize list, for-loop, or it.each iterates a hand-written subset of a dict/set/tuple/Enum that exists in the source, and names the members nothing covers. Use when reviewing tests, when a module…
gating
Build and audit deterministic verification gates — a check that blocks a pipeline and can be shown to go red. Use when writing a calibration gate, CI check, validation script or pre-publication check for a numeric or empirical result; when a plausible-but-wrong value would survive review; when asking whether an…
change-request-to-go-live-protocol
Use this skill when a Dynamics 365 change request must be structured and progressed through impact assessment, fit-gap analysis, UAT sign-off, go/no-go decision, go-live execution, and hypercare using the Success by Design framework. Defines the full change-to-go-live flow — change request intake, solution blueprint…
environment-to-production-release-protocol
Use this skill when a Power Platform or Dynamics 365 solution must progress through a structured dev-to-test-to-production release pipeline using managed solutions and Power Platform pipelines, when rollback readiness must be verified before go-live, or when a deployment approval gate must be enforced. Defines the…
erp-crm-cutover-protocol
Use this skill when a Dynamics 365 ERP or CRM implementation is approaching go-live and must progress through mock migration, data reconciliation, cutover runbook execution, rollback testing, and the Success by Design go-live gate before production switchover. Orchestrates d365-data-migration-cutover-agent as primary…