build-skill

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Super Mega Ultra Bestest Skill Builder

超级无敌最优Skill构建工具

Build skills by running 3 competing approaches in parallel, then merging the best of each.

通过并行运行3种不同的构建方案，再合并各方案的最优产出，来构建skill。

Args

参数

The args string is the skill specification. It should contain:

The skill name (required)
What the skill does, its modes/subcommands, and any context needed

Example:

/super-mega-ultra-bestest-skill-builder audit-permissions — wraps a TypeScript analyzer, default mode runs report, reset mode archives logs

If args are vague, ask one clarifying question before proceeding. Don't over-interview.

参数字符串为skill的规范说明，需要包含：

skill名称（必填）
skill的功能、模式/子命令，以及所有必要的上下文信息

示例：

/super-mega-ultra-bestest-skill-builder audit-permissions — wraps a TypeScript analyzer, default mode runs report, reset mode archives logs

如果参数表述模糊，在开始处理前先提1个澄清问题即可，不要过多询问。

Process

流程

1. Parse the spec

1. 解析规范

Extract from args:

Skill name (first word or hyphenated phrase before any separator)
Skill purpose (everything else)
Scope: user-level (
```
~/.claude/skills/
```
) or project-level (
```
.claude/skills/
```
) — default user-level unless the spec mentions a specific project

从参数中提取以下信息：

Skill名称：任意分隔符前的第一个单词或连字符连接的短语
Skill用途：其余所有内容
作用范围：用户级 (
```
~/.claude/skills/
```
) 或项目级 (
```
.claude/skills/
```
) — 除非规范中明确提到指定项目，否则默认是用户级

2. Launch 3 parallel agents

2. 启动3个并行Agent

Create

/tmp/skill-compare/{native,superpowers,manual}/

directories, then launch 3 background agents simultaneously via the Agent tool. Each gets the SAME spec but a DIFFERENT approach:

Agent 1 — "native-builder": Invoke

skill-creator:skill-creator

via the Skill tool, then follow its process. Write to

/tmp/skill-compare/native/SKILL.md

Agent 2 — "superpowers-builder": Invoke

superpowers:writing-skills

via the Skill tool, then follow its structural guidance (skip live subagent pressure testing but follow CSO, token efficiency, frontmatter, and checklist). Write to

/tmp/skill-compare/superpowers/SKILL.md

Agent 3 — "manual-builder": No skill-building guide. Write the SKILL.md using general best practices and intuition only. Write to

/tmp/skill-compare/manual/SKILL.md

All agents must be told:

Write ONLY to their
```
/tmp/skill-compare/<approach>/SKILL.md
```
path
Do NOT write to
```
~/.claude/skills/
```
or
```
.claude/skills/
```
The skill spec (passed through verbatim from args)
Brief context on what a Claude Code skill is (YAML frontmatter with
```
name
```
+
```
description
```
, markdown body with instructions)

创建

/tmp/skill-compare/{native,superpowers,manual}/

目录，然后通过Agent工具同时启动3个后台Agent。每个Agent拿到相同的规范说明，但采用不同的构建方案：

Agent 1 — "原生构建器"：通过Skill工具调用

skill-creator:skill-creator

，遵循其流程执行，结果写入

/tmp/skill-compare/native/SKILL.md

。

Agent 2 — "超能力构建器"：通过Skill工具调用

superpowers:writing-skills

，遵循其结构指引（跳过实例子代理压力测试，但要遵循CSO、token效率、前置元数据和检查清单要求），结果写入

/tmp/skill-compare/superpowers/SKILL.md

。

Agent 3 — "手动构建器"：不使用任何skill构建指南，仅基于通用最佳实践和经验编写SKILL.md，结果写入

/tmp/skill-compare/manual/SKILL.md

。

必须告知所有Agent：

仅可写入自己对应的
```
/tmp/skill-compare/<方案名>/SKILL.md
```
路径
不要写入
```
~/.claude/skills/
```
或
```
.claude/skills/
```
目录
完整传入参数中的skill规范说明
简单说明Claude Code skill的定义：包含
```
name
```
+
```
description
```
的YAML前置元数据，以及带操作指引的Markdown正文

3. Compare results

3. 对比结果

Once all 3 complete, read all 3 files and compare on these dimensions:

Dimension	What to evaluate
Discoverability	Does the description help Claude find it? Trigger-only (good) vs workflow summary (bad per CSO)?
Clarity	Can Claude follow instructions unambiguously? Are steps numbered?
Completeness	All modes covered? Edge cases? Troubleshooting?
Token efficiency	Word count vs information density. Target: <500 words for non-startup skills
Actionability	Concrete actions vs vague guidance? Explicit guardrails for failure modes?

Present the comparison as a table showing word counts, token costs, and per-dimension winners.

3个Agent全部执行完成后，读取3份结果文件，从以下维度对比：

维度	评估标准
可被检索性	描述是否能帮助Claude找到该skill？是否仅包含触发词（优）还是包含工作流总结（不符合CSO要求，劣）？
清晰度	Claude是否能无歧义地遵循指引？步骤是否有编号？
完整度	是否覆盖了所有模式？有没有考虑边缘场景？是否包含故障排查说明？
Token效率	字数和信息密度的比值，非启动类skill目标字数：<500字
可执行性	是具体可落地的操作还是模糊的指引？是否明确列出故障场景的防护规则？

输出对比表格，展示字数、token成本，以及每个维度的最优方案。

4. Synthesize

4. 整合生成

Cherry-pick the best elements from each approach into a final skill. For each element kept, note which approach it came from and why. Write the final skill to the target location:

User-level:
```
~/.claude/skills/<name>/SKILL.md
```
Project-level:
```
.claude/skills/<name>/SKILL.md
```

从每个方案中挑选最优元素，合并为最终的skill。每个保留的元素都要标注来源和选用原因。将最终skill写入目标路径：

用户级：
```
~/.claude/skills/<名称>/SKILL.md
```
项目级：
```
.claude/skills/<名称>/SKILL.md
```

5. Verify

5. 验证

Confirm the skill appears in the available skills list (check system reminder)
Report final word count
Show what was taken from each approach

确认skill出现在可用skill列表中（检查系统提醒）
报告最终字数
说明每个方案贡献了哪些内容

Known Patterns from Prior Runs

过往运行的已知规律

These patterns consistently emerge — use them to inform the merge:

Superpowers excels at: merge guardrails, CSO-compliant descriptions, cross-surface compatibility, explicit failure-mode prevention
Manual excels at: unique safety guardrails humans think of, natural "done" summary steps, concise structure
Native (skill-creator) excels at: comprehensive coverage, but tends to over-explain internals Claude doesn't need — trim aggressively
Your own judgment matters most for: token efficiency and cutting bloat

这些规律在过往运行中反复出现，可用于指导合并操作：

Superpowers 擅长：合并防护规则、符合CSO要求的描述、跨平台兼容性、明确的故障模式预防
手动方案 擅长：人类能想到的独特安全防护规则、自然的"完成"总结步骤、简洁的结构
原生（skill-creator） 擅长：全面的功能覆盖，但容易过度解释Claude不需要的内部逻辑 — 要大幅精简
你自己的判断 最适合用于：优化token效率、删减冗余内容