与子代理的环测试技巧

人工智能 社区

技能解读

当前原文的自动摘译,尚未人工审校。

何时使用

部署新技能之前 编辑现有技能之后 技能强制执行可以合理化的纪律

跳过时

纯参考技能 → 没有要测试的行为 没有代理有动机绕过的规则

相关

补充:ring:writing-skills、ring:test-driven-development 测试技能是应用于流程文档的 TDD。在没有技能的情况下运行场景(红色 - 观察代理失败),编写解决这些失败的技能(绿色),然后关闭漏洞(REFACTOR)。先决条件:首先了解ring:test-driven-development。完整的工作示例:examples/CLAUDE_MD_TESTING.md。

何时测试

测试技能:执行纪律(TDD、测试要求)、具有合规成本(时间、精力、返工)、可以合理化(“仅此一次”)或与近期目标相矛盾(速度高于质量)。跳过:纯参考技能(API文档),无规则可违反的技能。

TDD映射

TDD 阶段 · 技能测试 · 你做什么 红色 · 基线测试 · 在没有技能的情况下运行场景,观察代理失败 验证红色 · 捕获合理化 · 逐字记录确切的失败 绿色 · 编写技能 · 解决特定的基线故障 验证 绿色 · 压力测试 · 有技能运行,验证压力下的合规性

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
人工智能
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
88 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@LerianStudio · 未声明许可证
上下文用量评级
低消耗
接入复杂程度
即装即用
是否需要外部接口密钥
不需要
兼容的系统
未声明(默认跨平台)
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
  • 允许写入 / 修改
网络行为
仅限本地
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# Testing Skills With Subagents

## When to use
- Before deploying a new skill
- After editing an existing skill
- Skill enforces discipline that could be rationalized away

## Skip when
- Pure reference skill → no behavior to test
- No rules that agents have incentive to bypass

## Related
**Complementary:** ring:writing-skills, ring:test-driven-development

**Testing skills is TDD applied to process documentation.**

Run scenarios without the skill (RED — watch agent fail), write skill addressing those failures (GREEN), then close loopholes (REFACTOR).

**Prerequisite:** Understand `ring:test-driven-development` first. Complete worked example: `examples/CLAUDE_MD_TESTING.md`.

## When to Test

Test skills that: enforce discipline (TDD, testing requirements), have compliance costs (time, effort, rework), could be rationalized away ("just this once"), or contradict immediate goals (speed over quality).

**Skip:** Pure reference skills (API docs), skills without rules to violate.

## TDD Mapping

| TDD Phase | Skill Testing | What You Do |
|-----------|---------------|-------------|
| RED | Baseline test | Run scenario WITHOUT skill, watch agent fail |
| Verify RED | Capture rationalizations | Document exact failures verbatim |
| GREEN | Write skill | Address specific baseline failures |
| Verify GREEN | Pressure test | Run WITH skill, verify compliance under pressure |
| REFACTOR | Plug holes | Find new rationalizations, add counters |

## RED Phase: Watch It Fail

Run 3+ combined-pressure scenarios WITHOUT the skill. Document agent choices and rationalizations **word-for-word**.

**Why verbatim?** Exact wording reveals the loopholes to close.

### Writing Pressure Scenarios

| Quality | Example |
|---------|---------|
| Bad | "What does the skill say?" — agent recites |
| Good | "Production down, $10k/min, 5min window" — single pressure |
| Great | "3hr/200 lines done, 6pm, dinner plans, forgot TDD. A) Delete B) Commit C) Tests now" — multi-pressure + forced choice |

**Pressure types:** Time (deadline), sunk cost (hours invested), authority (senior says skip), economic (job at stake), exhaustion (end of day), pragmatic ("being realistic").

**Best tests combine 3+ pressures.**

## GREEN Phase: Write Minimal Skill

Address the specific failures documented in RED. Don't add hypothetical content — write just enough to address actual observed failures. Re-run same scenarios WITH skill; agent should now comply.

## REFACTOR Phase: Close Loopholes

Agent still violated rule despite having the skill? Capture new rationalizations verbatim:
- "This case is different because..."
- "I'm following the spirit not the letter"
- "Being pragmatic means adapting"

For each rationalization, add: explicit negation rule, rationalization table entry, red flag entry.

**Meta-test:** "You read the skill and chose wrong anyway. How could the skill have been written to make the right answer the only acceptable one?"

**Continue REFACTOR until no new rationalizations appear.**

## Signs of Bulletproof Skill

- Agent chooses correct option under maximum pressure
- Agent cites skill sections as justification
- Agent acknowledges temptation but follows rule anyway
- Meta-test reveals "skill was clear, I should follow it"

## Real-World Impact

From applying TDD to TDD skill itself:
- 6 RED-GREEN-REFACTOR iterations to bulletproof
- 10+ unique rationalizations discovered
- Each REFACTOR closed specific loopholes
- Final: 100% compliance under maximum pressure

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。