Evaluation Methodology
Use when measuring WorkBuddy Skill quality, interpreting evaluation results, calibrating thresholds, or deciding which triggering, output, robustness, or orchestration gap to improve first.
Install
npx skills add sandbaseai/workbuddy-skill --skill evaluation-methodologygit clone --depth 1 https://github.com/sandbaseai/workbuddy-skill.git /tmp/workbuddy-skill
mkdir -p ~/.claude/skills
cp -r /tmp/workbuddy-skill/skills/evaluation-methodology ~/.claude/skills/evaluation-methodologyPersonal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.
From the SKILL.md
# Skill 评估方法
评估 Skill 不只是数行数或看一条示例。将静态结构、受控评审和真实/模拟任务结果分层,分别回答“能否被正确触发”“是否能产出好结果”“是否稳定且适合编排”。每个分数都要能追溯到检查、样本和已知限制。
## 三层评估
### 1. 静态分析
快速、确定性、无模型调用。检查:
- frontmatter 是否包含清晰触发条件、名称和足够具体的描述;
- 输入/输出、代码示例、故障处理和相关 Skill 是否写清楚;
- 主文件是否精炼,细节是否合理分层;
- MUST/ALWAYS/NEVER 的密度和重复段落是否造成过度约束;
- 交叉引用和本地文件是否真实存在;
- 安全信号、秘密样式、悬空链接和未解释脚本是否被标记。
静态通过只证明结构检查通过,不能推断真实任务质量。静态发现的未知、超大文件或安全标记应进入报告,不得静默当成零风险。
### 2. 受控评审
使用至少一名对任务和评分锚点都不了解实现细节的评审者,分别给出 0–1 分:
| 维度 | 核心问题 |
| --- | --- |
| 触发准确性 | 相关请求会触发,不相关请求不会误触发吗? |
| 编排适配度 | Skill 是专注工作的 worker,还是越权充当 supervisor? |
| 输出质量 | 模拟真实任务时,结果完整、可执行、证据边界清楚吗? |
| 范围校准 | 既不是空泛短文,也没有把所有边缘情况塞进主上下文吗? |
评审输入应固定到 Skill 版本和同一份提示集;多个评审者时平均分,并记录分歧,而不是只保留最有利分数。
### 3. 任务模拟
在隔离环境运行一组真实但无副作用的代表性提示,记录:
- 触发率:相关样本中正确激活的比例;
- 输出一致性:质量分数的离散程度,不能只看最好的一次;
- 失败率:崩溃、违反格式、越权执行或证据缺失的比例;
- 成本效率:中位 token、IQR 和异常值;
- 是否产生未授权写入、网络调用、凭据暴露或不可逆副作用。
样本数量、提示来源、模型/运行时版本和随机种子要写入报告。模拟不能替代真实用户验证,也不能为了提高分数删除困难样本。
## 复合分数
除非项目已有固定权重,否则使用透明的默认权重,并在报告中披露:
| 维度 | 权重 |
| --- | ---: |
| 触发准确性 | 0.25 |
| 编排适配度 | 0.20 |
| 输出质量 | 0.15 |
| 范围校准 | 0.12 |
| 渐进披露 | 0.10 |
| token 效率 | 0.06 |
| 鲁棒性 | 0.05 |
| 结构完整性 | 0.03 |
| 模板质量 | 0.02 |
| 生态一致性 | 0.02 |
对每个维度,仅合并实际运行的层,并对可用层重新归一化:
```text
blended[d] = sum(weight[d, layer] * score[d, layer])
/ sum(weight[d, layer] for available layers)
composite = sum(dimension_weight[d] * blended[d]) * 100 * penalty
```
如果静态检查发现反模式,可使用 `penalty = max(0.5, 1 - 0.05 * count)`,并列出每个扣分原因。没有运行 judge 或模拟时,不得暗示复合分数代表完整认证。
## 分数解释与改进顺序Details
- Skill name
evaluation-methodology- Repository
- sandbaseai/workbuddy-skill
- Path
skills/evaluation-methodology/SKILL.md- License
- MIT
- Repo updated
- 8 September 2026
FAQ
What does the Evaluation Methodology skill do?
Use when measuring WorkBuddy Skill quality, interpreting evaluation results, calibrating thresholds, or deciding which triggering, output, robustness, or orchestration gap to improve first.
How do I install the Evaluation Methodology skill?
skills CLI (Claude Code, Cursor, Codex and more): npx skills add sandbaseai/workbuddy-skill --skill evaluation-methodology. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.
Is Evaluation Methodology popular?
It lives in sandbaseai/workbuddy-skill, which has 9 GitHub stars and 3 forks.
Related skills
Access Control Security Best Practices
9sandbaseai/workbuddy-skill
Design, review, implement, and verify least-privilege access controls across users, services, tenants, resources, and administrative workflows.
Accessibility Review
9sandbaseai/workbuddy-skill
Use when reviewing a web interface or interaction for semantic, keyboard, focus, screen-reader, ARIA, and contrast accessibility risks.
Acquire Codebase Knowledge
9sandbaseai/workbuddy-skill
Use when the user asks to map, document, or onboard into an existing codebase and needs evidence-backed architecture, structure, conventions, integrations, testing, and concerns documentation.
Evidence Before Claims
9sandbaseai/workbuddy-skill
Use immediately before claiming work is complete, fixed, passing, committed, published, or ready for handoff; require fresh verification evidence matched to the claim.
Distributed Tracing
9sandbaseai/workbuddy-skill
Use when instrumenting or diagnosing distributed requests, service dependencies, latency, retries, failures, or trace-to-log correlation.
Incident Postmortem
9sandbaseai/workbuddy-skill
Use after a production outage, significant degradation, data incident, or near miss to reconstruct evidence, quantify impact, identify systemic causes without blame, and create owned, dated prevention actions.
Browse the most popular Claude skills, all other skills, or write your own with the skills generator.