fusion-discover-skills

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Discover Fusion Skills

发现Fusion技能

When to use

适用场景

Use this skill when a user needs to discover which Fusion skill fits a task or wants MCP-backed guidance for installing, updating, or removing a skill.

Typical triggers:

"find a skill for this"
"what Fusion skill should I use?"
"is there a skill for GitHub issue authoring?"
"how do I install the right Fusion skill?"
"how do I update or remove installed Fusion skills?"
"what skills are available for this workflow?"

当用户需要发现适合任务的Fusion技能，或需要基于MCP的技能安装、更新或移除指导时，使用此技能。

典型触发场景：

"为这个任务找一个技能"
"我应该使用哪个Fusion技能？"
"有没有用于GitHub issue撰写的技能？"
"我该如何安装合适的Fusion技能？"
"我该如何更新或移除已安装的Fusion技能？"
"这个工作流有哪些可用技能？"

When not to use

不适用场景

Do not use this skill for:

creating or editing skills in a repository
doing the underlying task directly instead of helping with discovery
guessing skill matches when discovery data is unavailable
automatically installing, updating, or removing skills without the user's request

请勿将此技能用于：

在仓库中创建或编辑技能
直接执行底层任务而非提供发现帮助
在无发现数据时猜测技能匹配结果
未经过用户请求自动安装、更新或移除技能

Required inputs

必填输入项

Collect before responding:

the user's goal or query in their own words
the intended action:
```
query
```
,
```
install
```
,
```
update
```
, or
```
remove
```
the kind of skill the user is looking for: domain, workflow, or desired outcome
the active agent/client name when install guidance is needed and not already obvious
any constraints that materially change the recommendation, such as target runtime or repository context

If key inputs are missing, ask only the smallest question needed to resolve the intent. Use

references/follow-up-questions.md

for clarifying vague requests about which skill or workflow the user wants.

回复前需收集：

用户用自己的话描述的目标或查询内容
预期操作：
```
query
```
、
```
install
```
、
```
update
```
或
```
remove
```
用户寻找的技能类型：领域、工作流或期望结果
当需要安装指导且当前未明确时，获取活跃代理/客户端名称
任何会实质性改变推荐结果的约束条件，如目标运行时或仓库上下文

若关键输入缺失，仅询问解决意图所需的最简短问题。对于用户关于所需技能或工作流的模糊请求，使用

references/follow-up-questions.md

中的内容进行澄清。

Instructions

操作说明

If subagents are available, use the bundled advisor agents for specific discovery scenarios:

```
agents/fusion-mcp-advisor.md
```
when Fusion MCP is available — this is the primary discovery path
```
agents/github-mcp-advisor.md
```
when Fusion MCP is unavailable but GitHub MCP search is available
```
agents/github-raw-search-advisor.md
```
as a final fallback using read-only CLI and GraphQL inspection

If the runtime does not support bundled agents, follow the same workflow inline as described below.

Main workflow:

Detect the user's intent.
- Treat broad "what skill fits this?" requests as
```
query
```
  .
- Treat "install/add", "update/check updates", and "remove/uninstall" requests as explicit lifecycle intent.
- If the request mixes discovery and lifecycle steps, prioritize discovery first, then include the lifecycle guidance that follows from the selected skill.
Ask follow-up questions when the target skill is still unclear.
- If the request is vague, ask a minimal clarifying question about the task, domain, or workflow they want help with.
- Prefer questions that narrow the search space quickly: what they are trying to do, whether they want to discover, install, update, or remove a skill, and whether they already know a likely area such as GitHub workflow, MCP setup, or skill authoring.
- Use
```
references/follow-up-questions.md
```
  as the default question bank.
Ask for missing agent context only when needed.
- If the user wants install guidance and the active agent is not clear, ask which agent/client the command should target.
- Do not ask for agent details for plain discovery requests.
Query Fusion MCP first when available.
- Call
```
mcp_fusion_skills
```
  with the user's wording.
- Use auto intent detection by default, but set
```
intent
```
  explicitly when the user clearly wants
```
install
```
  ,
```
update
```
  ,
```
remove
```
  , or
```
query
```
  .
- Pass
```
agent
```
  for install intent when known so the advisory command is directly usable.
- Start with a small ranked set, usually top 3 to top 5 results.
Fall back to GitHub-backed discovery when Fusion MCP is unavailable or too weak.
- Prefer GitHub MCP repository or search tools when they are available in the current client.
- Otherwise use read-only shell-based inspection against trusted sources only, such as
```
npx skills add --list <source>
```
  , local
```
skills/**/SKILL.md
```
  searches,
```
gh search code
```
  , or
```
gh api graphql
```
  against the catalog repository.
- Use GraphQL only when structured repository data is needed and the higher-level MCP or search tools do not expose it cleanly.
- GraphQL read queries cost at least 1 point; keep
```
first
```
  /
```
last
```
  small (≤ 30) and avoid nested connections to minimize point cost. Do not retry on rate-limit errors; surface the failure and suggest retrying later.
- Budget awareness: a typical skill-discovery session should need at most 2–3 GitHub API calls (one search + one or two content reads). If the first search returns weak results, do one refinement pass and stop — do not loop.
- Treat GitHub-derived lifecycle guidance as fallback guidance unless Fusion MCP returned an explicit advisory command.
- Never use remote-script execution patterns or shell pipelines that execute fetched content.
Do one refinement pass when the first result set is weak.
- Rephrase the query with clearer domain keywords from the user's request.
- Stop after one refinement pass; do not keep searching indefinitely.
Return actionable results.
- For each recommended skill, include:
  - skill name
  - one-sentence purpose
  - why it matches the user's task
  - the next best action
- When MCP returns lifecycle guidance, relay the advisory command or instruction exactly, including any placeholder that still needs user input.
- When fallback discovery is used instead, label the source clearly as GitHub MCP, shell listing, code search, or GraphQL-derived guidance.
- When a skill is selected, include enough usage guidance that the user can proceed without another discovery round.
Handle low-confidence or no-match outcomes explicitly.
- Say that no strong match was found or that the recommendation is uncertain.
- If there are near matches, present them as tentative.
- Suggest the next best action: broaden the query, continue without a skill, or capture the gap for future skill authoring.
Keep the response scoped to discovery.
- Do not install, update, or remove anything unless the user explicitly asks you to execute that step.
- If Fusion MCP is unavailable, say so clearly and identify which GitHub-backed fallback path was used instead of inventing catalog results.

若有子代理可用，针对特定发现场景使用内置的顾问代理：

当Fusion MCP可用时，使用
```
agents/fusion-mcp-advisor.md
```
——这是主要的发现路径
当Fusion MCP不可用但GitHub MCP搜索可用时，使用
```
agents/github-mcp-advisor.md
```
当上述两者都不可用时，使用
```
agents/github-raw-search-advisor.md
```
作为最终回退方案，通过只读CLI和GraphQL检查

若运行时不支持内置代理，直接按照以下流程执行：

主工作流：

检测用户意图。
- 将宽泛的“哪个技能适合这个任务？”请求视为
```
query
```
  类型。
- 将“安装/添加”、“更新/检查更新”和“移除/卸载”请求视为明确的生命周期意图。
- 若请求同时包含发现和生命周期步骤，优先处理发现，然后提供所选技能对应的生命周期指导。
当目标技能不明确时，提出跟进问题。
- 若请求模糊，仅询问关于任务、领域或工作流的最简短澄清问题。
- 优先选择能快速缩小搜索范围的问题：用户想要完成什么、是想发现、安装、更新还是移除技能、是否已经知道可能的领域（如GitHub工作流、MCP设置或技能创作）。
- 默认使用
```
references/follow-up-questions.md
```
  中的问题库。
仅在需要时询问缺失的代理上下文。
- 若用户需要安装指导但活跃代理不明确，询问该命令应针对哪个代理/客户端。
- 对于纯发现请求，无需询问代理细节。
优先查询可用的Fusion MCP。
- 调用
```
mcp_fusion_skills
```
  并传入用户的表述。
- 默认使用自动意图检测，但当用户明确需要
```
install
```
  、
```
update
```
  、
```
remove
```
  或
```
query
```
  时，显式设置
```
intent
```
  参数。
- 当已知安装意图时，传入
```
agent
```
  参数，以便直接生成可用的指导命令。
- 先返回少量排序结果，通常为前3到前5个。
当Fusion MCP不可用或结果不佳时，回退到基于GitHub的发现方案。
- 若当前客户端支持，优先使用GitHub MCP仓库或搜索工具。
- 否则，仅对可信源使用基于shell的只读检查，如
```
npx skills add --list <source>
```
  、本地
```
skills/**/SKILL.md
```
  搜索、
```
gh search code
```
  或针对目录仓库的
```
gh api graphql
```
  查询。
- 仅当结构化仓库数据需要且高级MCP或搜索工具无法清晰暴露时，才使用GraphQL。
- GraphQL只读查询至少消耗1个积分；将
```
first
```
  /
```
last
```
  设置为较小值（≤30），避免嵌套连接以最小化积分消耗。遇到速率限制错误时不要重试；直接告知用户失败并建议稍后重试。
- 预算意识：一次典型的技能发现会话最多需要2-3次GitHub API调用（一次搜索+一到两次内容读取）。若首次搜索结果不佳，进行一次优化后停止——不要循环搜索。
- 除非Fusion MCP返回明确的指导命令，否则将GitHub衍生的生命周期指导视为回退方案。
- 绝不使用远程脚本执行模式或会执行获取内容的shell管道。
当首次结果集质量不佳时，进行一次优化。
- 使用用户请求中更清晰的领域关键词重新表述查询。
- 优化一次后停止；不要无限搜索。
返回可执行的结果。
- 对于每个推荐技能，包含：
  - 技能名称
  - 一句话描述用途
  - 为何匹配用户任务
  - 最佳下一步操作
- 当MCP返回生命周期指导时，准确传达指导命令或说明，包括仍需用户输入的占位符。
- 若使用回退发现方案，需明确标注来源为GitHub MCP、shell列表、代码搜索或GraphQL衍生指导。
- 当选定技能后，提供足够的使用指导，确保用户无需再次进行发现操作即可继续。
明确处理低置信度或无匹配结果的情况。
- 告知用户未找到强匹配结果或推荐存在不确定性。
- 若有近似匹配，将其作为暂定结果呈现。
- 建议最佳下一步操作：拓宽查询范围、不使用技能继续操作，或记录此缺口以便未来开发新技能。
保持响应范围仅限于发现环节。
- 除非用户明确要求执行该步骤，否则不要安装、更新或移除任何内容。
- 若Fusion MCP不可用，明确告知用户，并说明使用了哪个基于GitHub的回退方案，而非编造目录结果。

Examples

示例

User: "Find a Fusion skill for GitHub issue authoring."
- Result: return
```
fusion-issue-authoring
```
  , explain that it routes to issue-type-specific skills, and include the install guidance or next step returned by MCP.
User: "How do I update my installed Fusion skills?"
- Result: call
```
mcp_fusion_skills
```
  with
```
intent: update
```
  and return the advisory update command or instructions from MCP.
User: "Is there a skill for this obscure workflow?"
- Result: if Fusion MCP is unavailable or yields weak matches, use GitHub-backed fallback discovery, say whether the match is tentative, and suggest the next best action.

用户：“找一个用于GitHub issue撰写的Fusion技能。”
- 结果：返回
```
fusion-issue-authoring
```
  ，说明它会路由到特定issue类型的技能，并包含MCP返回的安装指导或下一步操作。
用户：“我该如何更新已安装的Fusion技能？”
- 结果：调用
```
mcp_fusion_skills
```
  并设置
```
intent: update
```
  ，返回MCP提供的更新指导命令或说明。
用户：“有没有适合这个小众工作流的技能？”
- 结果：若Fusion MCP不可用或匹配结果不佳，使用基于GitHub的回退发现方案，说明匹配结果是否为暂定，并建议下一步操作。

Expected output

预期输出

Return:

detected intent (
```
query
```
,
```
install
```
,
```
update
```
, or
```
remove
```
)
source used for the recommendation (
```
mcp_fusion_skills
```
, GitHub MCP, shell listing, code search, or GraphQL fallback)
ranked skill suggestions when available
concise description of each recommended skill
next-step guidance for the chosen result
advisory install/update/remove command when MCP provides one
clear uncertainty language and a fallback next action when no strong match exists

返回内容应包含：

检测到的意图（
```
query
```
、
```
install
```
、
```
update
```
或
```
remove
```
）
推荐结果的来源（
```
mcp_fusion_skills
```
、GitHub MCP、shell列表、代码搜索或GraphQL回退）
可用时返回排序后的技能建议
每个推荐技能的简洁描述
所选结果的下一步指导
当MCP提供时，返回安装/更新/移除的指导命令
当无强匹配结果时，使用清晰的不确定性表述和回退下一步操作

Agents

子代理

Helper agents for specific discovery scenarios (when subagents are available):

agents/fusion-mcp-advisor.md — use when Fusion MCP is available
agents/github-mcp-advisor.md — use when Fusion MCP is unavailable but GitHub MCP is available
agents/github-raw-search-advisor.md — use as final fallback with read-only CLI/GraphQL

针对特定发现场景的辅助代理（当子代理可用时）：

agents/fusion-mcp-advisor.md — 当Fusion MCP可用时使用
agents/github-mcp-advisor.md — 当Fusion MCP不可用但GitHub MCP可用时使用
agents/github-raw-search-advisor.md — 使用只读CLI/GraphQL作为最终回退方案

References

参考资料

references/follow-up-questions.md — clarifying questions for vague requests

references/follow-up-questions.md — 用于澄清模糊请求的问题库

Safety & constraints

安全与约束

Never:

invent skill names, availability, or install/update/remove commands
present low-confidence matches as certain
use remote-script execution patterns or unsafe shell pipelines during fallback discovery
mutate the user's environment unless they explicitly ask you to execute the next step

Always:

prefer Fusion MCP first, then GitHub MCP, then read-only shell or GraphQL fallback
keep suggestions concise and task-focused
preserve any advisory command text exactly when MCP returns it
label GitHub-derived fallback guidance clearly when MCP was not the source
state uncertainty plainly when the evidence is weak

绝对禁止：

编造技能名称、可用性或安装/更新/移除命令
将低置信度匹配结果表述为确定结果
在回退发现过程中使用远程脚本执行模式或不安全的shell管道
除非用户明确要求执行下一步，否则不要修改用户环境

必须遵守：

优先使用Fusion MCP，其次是GitHub MCP，最后是只读shell或GraphQL回退方案
保持建议简洁且聚焦任务
当MCP返回指导命令时，完全保留原文，包括占位符
当未使用MCP时，明确标注基于GitHub的回退指导来源
当证据不足时，清晰说明不确定性