npx skills add https://github.com/equinor/fusion-skills --skill fusion-skill-authoring当您需要在 skills/ 目录下创建新技能,或现有技能需要进行内容创作层面的更新而非简单的文本编辑时,请使用此技能。
典型触发场景:
skills/<名称>/SKILL.md 的脚手架"隐含触发场景:
以下情况请勿使用此技能:
skills/ 之外的产品或应用程序代码如果缺少必需输入,请先提出简洁、有针对性的问题。使用 assets/follow-up-questions.md 作为默认问题库。
起草前收集:
Use this skill when you need to create a new skill under skills/, or when an existing skill needs a material authoring refresh instead of a small copy edit.
Typical triggers:
skills/<name>/SKILL.md"Implicit triggers:
Do not use this skill for:
skills/广告位招租
在这里展示您的产品或服务
触达数万 AI 开发者,精准高效
custom-<基础技能名称>相关时收集:
metadata.orchestrator,编排器的 metadata.skillsactive、experimental、deprecated、archived)compatibility 文本metadata.mcp.required / metadata.mcp.suggested)agents/ 辅助集是否能优化范围界定、审查或触发调优scripts/ 目录是合理的有用时捕获:
metadata.sponsor 作为备用责任方写入文件前验证:
name:1-64个字符,仅限小写字母/数字/连字符,必须与文件夹名称匹配,无前导/尾随连字符,无连续连字符,无 XML 标签,且无保留字custom-<基础技能名称>description:非空,<= 1024 字符,第三人称,无 XML 标签,说明技能的作用以及何时使用它
USE FOR: 和 DO NOT USE FOR: 线索的单引号 YAML 字符串description: '根据已验证的仓库上下文起草发布说明。USE FOR: 发布摘要、更新日志准备。DO NOT USE FOR: 发布发布或编辑产品代码。'metadata.version:遵循目标目录的起始版本规则;如果无本地规则,"0.0.0" 是新技能的安全默认值metadata.owner:仅当目标目录要求明确所有权时包含;使用稳定的 GitHub 身份或等效团队句柄metadata.status:仅当目标目录跟踪生命周期状态时包含;如果使用,请保持为 active、experimental、deprecated 或 archivedmetadata.tags:保持标签相关、小写和 kebab-casemetadata:除非关系字段明确需要列表或映射,否则使用简单的键/值元数据license:可选的顶级字段compatibility:可选的顶级字段;仅当技能有实际的运行时、网络、工具或产品约束时才包含仓库特定的前缀规则、所有权/生命周期要求、发布策略和验证命令属于仓库本地指令或目录文档,不属于可移植的技能包。
在编写长篇指令前,至少捕获三个代表性请求:
将这些请求作为最终技能的验收标准。如果无法定义现实的请求,则范围可能未充分指定或可复用性不足以成为技能。
预先决定技能类型:
能力提升:打包智能体尚不具备的领域知识、工具或参考资料工作流程/编码偏好:打包用户希望一致重复的序列、审查门控、样式规则或变更顺序决定组合模式:
独立:无需协调技能编排器:路由到配套技能并拥有共享门控从属:仅在其编排器下运行,并应清晰记录该依赖关系选择支持任务的最小文件夹结构:
SKILL.mdreferences/ 用于长篇指南、示例、表格或平台特定细节assets/ 用于模板、检查清单、示例输出和静态文件agents/ 用于少量专门的辅助角色,当运行时支持技能本地智能体且第二轮能实质性改进范围界定、审查或触发质量时scripts/ 仅当确定性自动化能实质性提高安全性或可靠性时保持参考资料从 SKILL.md 开始仅一级深度。不要创建强制部分阅读的嵌套引用链。
SKILL.md首先编写最小可用的主文档:
何时使用 和 何时不使用必需输入指令预期输出安全与约束将草稿保持在 300 行以内。不同的智能体运行时具有不同的上下文窗口限制——超过 300 行的 SKILL.md 文件在较小窗口的运行时上存在性能下降风险并会触发 CI 警告。超过 500 行的文件会导致 CI 失败。尽早将溢出内容移至 references/,而不是后期裁剪。
优先使用简洁、具体的指令,而非背景解释。假设智能体已具备能力,仅添加其无法可靠推断的上下文。
有意设定自由度:
在 SKILL.md 中至少包含一个具体示例,或直接链接到 references/ 中的一个。
当长篇或专门内容能提高清晰度时,将其移出 SKILL.md:
references/ 存放深度指南、大型示例、API/平台说明或长篇检查清单assets/ 存放技能应直接指向的模板和可复用工件agents/ 存放少量辅助集,当运行时支持技能本地智能体或子智能体且工作流程受益于范围界定的第二意见时scripts/ 仅用于应执行而非重新生成的确定性操作如果添加脚本:
如果运行时忽略捆绑的辅助智能体,请内联遵循相同的角色,而不是跳过评估步骤。
如果技能依赖 MCP,请在 metadata.mcp 中声明要求,并在技能内容中记录客户端特定的工具命名期望,而不是假设所有运行时的行为方式相同。
在创作更改后,运行目标环境支持的验证:
如果环境没有专门的技能工具:
SKILL.md 和每个直接引用的文件使用步骤 2 中的代表性请求来审查最终结果:
如果子智能体可用,请在它们有帮助时使用捆绑的角色文件:
agents/scoper.md 来决定创建、更新还是非技能,并选择最小有用的结构agents/reviewer.md 来审查发现、结构、安全性和验证证据,如同严格的维护者agents/trigger-tuner.md在可移植包正确后,应用仓库本地指令或等效目录策略中的任何仓库特定发布或版本控制规则。
将创作结果作为明确的契约返回:
SKILL.md此技能借鉴了 Anthropic skill-creator 捆绑小型 agents/ 辅助集的模式,但将角色缩小到 Fusion 特定的范围界定、审查和触发调优。
agents/scoper.md — 决定请求应成为新技能、更新还是非技能;选择仍能解决问题的最小文件夹结构agents/reviewer.md — 根据发现、结构、安全性和验证期望审查已起草的技能包agents/trigger-tuner.md — 锐化描述措辞,并针对现实提示比较激活线索变体如果运行时不提供子智能体,请保持相同的审查循环内联,不要仅仅因为打包被忽略而跳过智能体形态的推理。
SKILL.md,仅在需要模板时添加 assets/。fusion-skill-authoring 的激活线索和结构。"
返回:
有关默认文件夹结构和 SKILL.md 基线模板,请参阅 references/skill-template-baseline.md。
有关成功信号、常见失败信号和恢复步骤,请参阅 references/validation-signals.md。
使用 assets/skill-readiness-checklist.md 作为技能更改的最终质量检查清单。仓库特定的 PR 要求属于仓库指令,不属于可安装技能资产。
切勿:
始终:
SKILL.md 简洁,并将溢出内容移至直接引用每周安装数
77
仓库
首次出现
2026年2月20日
安全审计
安装于
github-copilot77
opencode70
gemini-cli70
codex70
kimi-cli70
amp70
If required inputs are missing, ask concise targeted questions first. Use assets/follow-up-questions.md as the default question bank.
Collect before drafting:
custom-<base-skill-name> unless the target repository has a different naming conventionCollect when relevant:
metadata.orchestrator for subordinates, metadata.skills for orchestratorsactive, experimental, deprecated, archived) when the target catalog tracks lifecycle statecompatibility text when the skill has real environment constraintsmetadata.mcp.required / metadata.mcp.suggested) when the skill depends on specific serversagents/ helper set would sharpen scoping, review, or trigger tuningscripts/ directoryCapture if useful:
metadata.sponsor as backup accountability if the target catalog uses backup ownershipValidate before writing files:
name: 1-64 characters, lowercase letters/numbers/hyphens only, must match the folder name, no leading/trailing hyphen, no consecutive hyphens, no XML tags, and no reserved wordscustom-<base-skill-name>description: non-empty, <= 1024 chars, third-person, no XML tags, states both what the skill does and when to use it
USE FOR: and DO NOT USE FOR: cuesdescription: 'Drafts release notes from validated repository context. USE FOR: release summaries, changelog preparation. DO NOT USE FOR: publishing releases or editing product code.'metadata.version: follow the target catalog's starting-version rule; if no local rule exists, "0.0.0" is a safe default for a new skillmetadata.owner: include only when the target catalog requires explicit ownership; use a stable GitHub identity or equivalent team handlemetadata.status: include only when the target catalog tracks lifecycle state; if used, keep it to active, experimental, deprecated, or archivedmetadata.tags: keep tags relevant, lowercase, and kebab-casemetadata: use simple key/value metadata unless a relationship field explicitly needs a list or maplicense: optional top-level fieldcompatibility: optional top-level field; only include it when the skill has real runtime, network, tool, or product constraintsRepository-specific prefix rules, ownership/lifecycle requirements, release policy, and validation commands belong in repo-local instructions or catalog docs, not in the portable skill package.
Capture at least three representative requests before writing long instructions:
Use these requests as the acceptance criteria for the final skill. If you cannot define realistic requests, the scope is probably underspecified or not reusable enough to become a skill.
Decide the skill type up front:
capability uplift: packages domain knowledge, tools, or reference material the agent does not already haveworkflow / encoded preference: packages sequencing, review gates, style rules, or mutation order the user wants repeated consistentlyDecide the composition model:
standalone: no coordinating skill requiredorchestrator: routes to companion skills and owns shared gatessubordinate: runs only under its orchestrator and should document that dependency clearlyChoose the minimum folder structure that supports the task:
SKILL.md alwaysreferences/ for long guidance, examples, tables, or platform-specific detailsassets/ for templates, checklists, sample outputs, and static filesagents/ for a small number of specialized helper roles when the runtime supports skill-local agents and a second pass materially improves scoping, review, or trigger qualityscripts/ only when deterministic automation materially improves safety or reliabilityKeep references one level deep from SKILL.md. Do not create nested reference chains that force partial reads.
SKILL.mdWrite the smallest useful main document first:
When to use and When not to useRequired inputsInstructionsExpected outputSafety & constraintsKeep the draft under 300 lines. Different agent runtimes have different context window limits — SKILL.md files over 300 lines risk degradation on smaller-window runtimes and trigger CI warnings. Files over 500 lines fail CI. Move overflow to references/ early rather than trimming later.
Prefer concise, specific instructions over background explanation. Assume the agent is already capable and only add context it would not reliably infer.
Set the degree of freedom intentionally:
Include at least one concrete example in SKILL.md or link directly to one in references/.
Move long or specialized content out of SKILL.md when it improves clarity:
references/ for deep guidance, large examples, API/platform notes, or long checklistsassets/ for templates and reusable artifacts the skill should point at directlyagents/ for a small helper set when the runtime supports skill-local agents or subagents and the workflow benefits from a scoped second opinionscripts/ only for deterministic operations that should be executed instead of regeneratedIf you add scripts:
If the runtime ignores bundled helper agents, follow the same roles inline instead of skipping the evaluation step.
If the skill depends on MCP, declare the requirement in metadata.mcp and document client-specific tool naming expectations in the skill content instead of assuming all runtimes behave the same way.
Run the validation supported by the target environment after authoring changes:
If the environment has no dedicated skill tooling:
SKILL.md and every directly referenced file end-to-endUse the representative requests from Step 2 to review the final result:
If subagents are available, use the bundled role files when they help:
agents/scoper.md before drafting to decide create vs update vs not-a-skill and to choose the smallest useful structureagents/reviewer.md after drafting to review discovery, structure, safety, and validation evidence like a strict maintaineragents/trigger-tuner.md when the main risk is weak activation cues or when choosing between two description variantsAfter the portable package is correct, apply any repository-specific release or versioning rules from repo-local instructions or equivalent catalog policy.
Return the authoring result as an explicit contract:
SKILL.mdThis skill borrows Anthropic skill-creator's pattern of bundling a small agents/ helper set, but narrows the roles to Fusion-specific scoping, review, and trigger tuning.
agents/scoper.md — decide whether the request should become a new skill, an update, or not a skill at all; choose the smallest folder structure that still solves the problemagents/reviewer.md — review a drafted skill package against discovery, structure, safety, and validation expectationsagents/trigger-tuner.md — sharpen description wording and compare activation-cue variants against realistic promptsIf a runtime offers no subagents, keep the same review loop inline and do not skip the agent-shaped reasoning just because the packaging is ignored.
SKILL.md, and add assets/ only if templates are needed.fusion-skill-authoring."
Return:
See references/skill-template-baseline.md for the default folder structure and SKILL.md baseline template.
See references/validation-signals.md for success signals, common failure signals, and recovery steps.
Use assets/skill-readiness-checklist.md as the final-quality checklist for skill changes. Repository-specific PR requirements belong in repository instructions, not in the installable skill asset.
Never:
Always:
SKILL.md concise and move overflow to direct referencesWeekly Installs
77
Repository
First Seen
Feb 20, 2026
Security Audits
Installed on
github-copilot77
opencode70
gemini-cli70
codex70
kimi-cli70
amp70
AI Elements:基于shadcn/ui的AI原生应用组件库,快速构建对话界面
69,600 周安装
ISO 13485质量文档管理系统:编号、审批、变更控制与21 CFR Part 11合规指南
173 周安装
OAuth 2.1 迁移指南与安全最佳实践 | 认证安全专家详解
70 周安装
LatchBio集成指南:Python生物信息学无服务器工作流开发与部署
177 周安装
ISO 14971:2019 医疗器械风险管理专家指南 | 全生命周期风险分析、评估与控制
174 周安装
Capacitor 启动画面插件:iOS/Android 自定义启动页、动画与问题解决指南
175 周安装
Shopware 6 最佳实践指南:22类77条规则,助力电商插件开发与性能优化
177 周安装