自定义技能进阶

把反复使用的写作方法存为技能，减少每次重新解释。一个技能最好解决一项明确问题，写清依据、步骤、交付形式和边界。

从一次真实任务开始

先在对话中完成一次你满意的局部任务，再提炼方法。不要把某篇作品的人名和未公开正文直接写成通用技能。

下面的“有限视角检查”可以放入指令区：

# 有限视角检查

## 依据
先读取作者指定的人物资料、当前场景与引用段落。
缺少人物已知信息时，先提出问题。

## 步骤
1. 找出直接断言其他人物内心、身份或真相的句子。
2. 对照当前视角人物已经知道的事。
3. 区分确有矛盾与尚待作者决定的问题。

## 交付
列出问题原句、文件依据和最小替换建议。
保留已经确认的行动与结果。
先在对话中给建议，不修改文件，不补写人物生平。

技能提供写作方法；本轮所需资料仍要通过打开目标文件、引用原文或点名读取来提供。{{selectedText}}、{{currentFile}} 不会自动替换为项目内容。

触发词与主动选择

用容易记住、能说明任务的短语，例如“检查视角”“有限视角”。多个词用英文逗号分隔。避免给所有技能都设置“修改”这样宽泛的词。

想明确使用某技能，在 / 菜单主动选择。每轮最多选择 3 个，复杂任务可按阶段使用不同方法。

添加资源

保存技能后，可以维护资源文件，把较长模板和参考资料放在 references/ 或 assets/ 下。技能指令中说明何时读取哪些资源。

资源支持 Markdown、TXT、JSON、YAML 和 CSV 等文本格式。技能可以提供方法和参考文本，不执行包内脚本。

导入标准技能包

导入入口接受 Markdown 或 ZIP。标准包包含 SKILL.md，文件头至少有 name 和 description：

---
name: viewpoint-check
description: Check a scene against the viewpoint character's knowledge.
---

Read the specified scene and character notes.
List inconsistent sentences, evidence, and suggested replacements.

ZIP 包不超过 1MiB，可包含文本资源目录。导入后检查警告：脚本和不支持的二进制文件会被跳过，allowed-tools 不会改变工作台权限。

验证和迭代

用短段落检查方法是否解决了问题，再试一次资料缺失和一次事实冲突。观察问题定位与修改范围，而不是只看技能是否被选中。

不符合预期时先改最相关的一条要求。指令过长、多个方法互相冲突时，拆为不同技能。分享前导出自己的副本，移除私密或无权公开的内容。

基础操作见使用技能。

Advanced custom skills

Save methods you use repeatedly so you do not explain them every time. A skill works best on one clear problem with evidence, steps, a deliverable, and boundaries.

Start from a real task

Complete a useful local task in chat, then extract its method. Keep private manuscripts and story-specific names out of a general-purpose skill.

Use this limited-viewpoint method in the instructions field:

# Limited viewpoint check

## Evidence
Read the specified character notes, current scene, and quoted passage.
Ask if the character's knowledge is missing.

## Steps
1. Find assertions about another character's thoughts, identity, or hidden truth.
2. Compare them with what the viewpoint character already knows.
3. Separate contradictions from decisions still open to the author.

## Deliverable
List problem sentences, file evidence, and the smallest replacements.
Retain confirmed actions and outcomes.
Suggest changes in chat first. Do not edit files or invent biographies.

A skill provides a method. Supply this turn's material by opening the target, quoting text, or naming files to read. {{selectedText}} and {{currentFile}} are not automatically replaced with project content.

Trigger phrases and explicit selection

Use memorable, task-specific phrases such as “check viewpoint” and “limited viewpoint.” Separate them with English commas. Avoid broad words such as “edit” for every skill.

Select a skill explicitly through / when you want it used. Choose up to 3 per turn; use different methods at different stages for complex tasks.

Add resources

After saving a skill, maintain longer templates and reference text under references/ or assets/. Explain in the instructions when each resource should be read.

Resources support text formats including Markdown, TXT, JSON, YAML, and CSV. Skills provide methods and reference text; package scripts do not run.

Import a standard package

Import accepts Markdown or ZIP. A standard package contains SKILL.md with at least name and description in its header:

---
name: viewpoint-check
description: Check a scene against the viewpoint character's knowledge.
---

Read the specified scene and character notes.
List inconsistent sentences, evidence, and suggested replacements.

ZIP packages must be at most 1MiB and may include text-resource directories. Check import warnings: scripts and unsupported binaries are skipped, and allowed-tools does not change workbench permissions.

Check and improve

Try a short passage, then missing evidence and a conflicting fact. Inspect problem detection and edit scope, not just whether the skill was selected.

Change the most relevant instruction first. Split long or conflicting methods into separate skills. Before sharing, export your own copy and remove private or unlicensed text.

See using skills for basic operations.