技能发展

工程开发 社区 v0.2.0

技能解读

当前原文的自动摘译,尚未人工审校。

目标

产生这样的技能:易于触发,依赖于SKILL.md层,由真实的references/,examples/和scripts/文件支持,没有死的本地引用。

核心规则

保留一项技能=一份持久的工作。将前题描述视为主要触发面。让SKILL.md专注于工作流程和边界。将详细的目录、模板和冗长的说明移至references/或examples/。不要提及不存在的文件。

1.先检查当前环境

在编写任何内容之前:检查目标技能目录,检查已经解决类似问题的相邻技能,验证实际存在哪些代理、命令和同级技能,在添加新引用之前识别过时的引用。

2.锁定技能合约

在编辑之前定义四件事:技能的作用、触发它的内容、它明确不执行的操作、实际需要哪些捆绑资源。

3. 编写或修复frontmatter

前面的内容应该:使用name中的真实技能标识符,使用第三人称触发描述,包含用户自然会说的具体短语,保持足够短以便快速浏览。

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
工程开发
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
92 / 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
@Galaxy-Dawn · v0.2.0 · 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

# Skill Development

Use this skill to create or repair Claude skills in the **current local environment**, not in an abstract plugin template.

## Goal

Produce a skill that is:
- easy to trigger,
- lean at the `SKILL.md` layer,
- backed by real `references/`, `examples/`, and `scripts/` files when they are mentioned,
- free of dead local references.

## Core rules

- Keep **one skill = one durable job**.
- Treat the frontmatter description as the main trigger surface.
- Keep `SKILL.md` focused on workflow and boundaries.
- Move detailed catalogs, templates, and long explanations into `references/` or `examples/`.
- Do not mention files that do not exist.
- Do not inherit stale names, agents, or sibling skill references without verifying they exist locally.

## Default workflow

### 1. Inspect the current environment first

Before writing anything:
- inspect the target skill directory,
- inspect neighboring skills that already solve a similar problem,
- verify which agents, commands, and sibling skills actually exist,
- identify stale references before adding new ones.

Use the local inventory as the authority. Do not write guidance against an imagined plugin layout.

### 2. Lock the skill contract

Define four things before editing:
1. what the skill does,
2. what triggers it,
3. what it explicitly does **not** do,
4. which bundled resources are actually needed.

If the skill only needs a short workflow, keep it short. Do not create `references/`, `examples/`, or `scripts/` just because the directories are conventional.

### 3. Write or repair the frontmatter

The frontmatter should:
- use the real skill identifier in `name`,
- use a third-person trigger description,
- include concrete phrases a user would naturally say,
- stay short enough to scan quickly.

Prefer descriptions of this form:

```yaml
---
name: skill-name
description: This skill should be used when the user asks to "...", "...", or needs help with ....
---
```

### 4. Keep the main file lean

A good `SKILL.md` should usually contain:
- a short goal section,
- role boundaries,
- a default workflow,
- safety or quality rules,
- a short list of additional resources.

Move these out of the main file when they get long:
- templates,
- exhaustive checklists,
- edge-case catalogs,
- sample outputs,
- long examples.

### 5. Add only real bundled resources

Use bundled resources deliberately:
- `references/` for detailed guidance that may be loaded selectively,
- `examples/` for real example outputs or scaffolds,
- `scripts/` for deterministic helper logic.

If a resource is mentioned in `SKILL.md`, it must exist.
If a resource exists but is never referenced or used, delete it.

### 6. Run integrity checks before closing

At minimum, verify:
- frontmatter parses,
- referenced local files exist,
- sibling skill or agent references are real,
- `SKILL.md` is not overloaded with material that belongs in references,
- temporary logs, caches, and editor artifacts are not left inside the skill directory.

## Typical repair patterns

### When the skill is too long
- keep the trigger and workflow in `SKILL.md`,
- move catalogs and deep detail into `references/`,
- keep a short read order so another model knows what to load first.

### When the skill is too thin
- add a default workflow,
- add at least one concrete example or checklist,
- make the boundaries explicit so the skill is not just a slogan.

### When the skill has stale references
- remove dead paths immediately,
- replace historical names with current local names,
- re-check neighboring agents/commands/skills against the live directory.

## Recommended output shape

When creating or repairing a skill, prefer ending with:
- what changed,
- which files were created or updated,
- what integrity checks were run,
- what still needs manual follow-up, if anything.

## References

Load only what is needed:
- `references/checklist.md` - compact quality checklist before closing a skill edit
- `references/integrity-checks.md` - concrete local checks for missing files, dead references, and drift
- `references/skill-creator-original.md` - legacy background reference; use for context, not as the live source of truth

讨论

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