skillify

作者: microsoft

从当前会话中捕获可重复的流程,生成可复用的 Amplifier SKILL.md 技能文件。分析对话,询问用户以确认…

npx skills add https://github.com/microsoft/amplifier-bundle-skills --skill skillify

Skillify

Create a well-structured, reusable Amplifier SKILL.md skill file that captures a repeatable process from the current session so it can be invoked again later via /skill-name. The skill must conform to the Agent Skills specification with Amplifier extensions.

Inputs

  • $ARGUMENTS: (Optional) Description of the process to capture as a skill.

Steps

1. Consult Skills-Assist

Before writing any skill content, load the authoritative skills reference:

load_skill("skills-assist")

Ask skills-assist about:

  • Current frontmatter field reference and best practices
  • Fork vs inline decision criteria
  • Step annotation conventions
  • Testing patterns for the type of skill being created

skills-assist is the source of truth for Amplifier skill conventions. The examples in this skill are illustrative samples — consult skills-assist for the complete and up-to-date specification.

Success criteria: You have loaded and consulted skills-assist for the latest skill authoring conventions.

2. Analyze the Session

Before asking the user anything, analyze the conversation history to identify:

  • What repeatable process was performed
  • What the inputs and parameters were
  • The distinct steps in order
  • The success criteria and artifacts for each step (not just "writing code" but "an open PR with CI passing")
  • Where the user corrected or steered you — these are important design signals
  • What tools were used (read_file, write_file, bash, glob, grep, delegate, etc.)
  • What agents were delegated to
  • What the goals and measurable outcomes were

Success criteria: You have a clear mental model of the process, its steps, inputs, outputs, and success criteria.

3. Interview the User

Output density rule: Group related decisions into natural clusters — not one question per turn (tedious) and not everything at once (overwhelming). Present your analysis first, let the user absorb it, then ask related questions together. For example, present identity/routing decisions as one cluster, execution model decisions as another.

Calibrate interview depth to the complexity of the process. A simple 2-step workflow needs 1-2 rounds. A complex multi-step workflow with parallel tasks and irreversible actions needs the full treatment.

Round 1: High-level confirmation

  • Suggest a name and description for the skill based on your analysis. Ask the user to confirm or rename.
  • Suggest high-level goals and specific success criteria.
  • Present the high-level steps you identified as a numbered list.

Round 2: Details and arguments

  • If the skill needs arguments, suggest them based on what you observed.
  • Ask whether this skill should run inline (in the current conversation) or forked (context: fork) as an isolated subagent. Forked is better for self-contained tasks that don't need mid-process user input; inline is better when the user wants to steer mid-process.
  • Ask where the skill should be saved:
    • This project (.amplifier/skills/<name>/SKILL.md) — project-specific
    • Personal (~/.amplifier/skills/<name>/SKILL.md) — follows you across projects
    • The skills bundle (amplifier-bundle-skills/skills/<name>/SKILL.md) — if contributing to the curated collection

Round 3: Per-step breakdown (complex processes only)

Skip this round for simple skills with obvious steps. For complex skills:

  • What does each step produce that later steps need?
  • What proves each step succeeded?
  • Should the user confirm before irreversible actions?
  • Are any steps independent and could run in parallel?
  • How should each step be executed? (direct, delegate to an agent, user action)
  • What are hard constraints or preferences?

Pay special attention to places where the user corrected you during the session. These corrections are the most valuable design signals.

Round 4: Final confirmation

  • Confirm when this skill should be invoked and review trigger phrases for the description field.
  • Ask for gotchas or edge cases, if still unclear.

Stop interviewing once you have enough information. Don't over-ask for simple processes.

Success criteria: You have all the information needed to write the SKILL.md and the user has confirmed the design.

4. Write the SKILL.md

Create the skill directory and file at the location the user chose in Round 2.

Use this template as a starting point. Consult skills-assist for the full set of available frontmatter fields and current conventions:

---
name: {{skill-name}}
description: >
  {{TRIGGER FIRST: the condition under which this applies, then what it
  does. Then "USE WHEN <deciding factor>." Then "DO NOT USE WHEN <the
  case that belongs elsewhere> — use <name>." This is what the model sees
  in the skills visibility list, on EVERY request of every session that
  can see the skill, invoked or not.
  HARD CAP 400 characters (ERROR at 800) — enforced by
  foundation:recipes/validate-bundle-repo.yaml Phase 2.82.
  ZERO <example> blocks and zero <commentary> tags: a trigger belongs in
  the USE WHEN clause as a decision rule, not as a worked dialogue.}}
user-invocable: true
allowed-tools:
  {{list of Amplifier tool names observed during the session, e.g.:}}
  {{- read_file}}
  {{- write_file}}
  {{- edit_file}}
  {{- bash}}
  {{- glob}}
  {{- grep}}
  {{- delegate}}
model_role: {{general | coding | reasoning | critique | writing | fast}}
{{Only include if forked:}}
{{context: fork}}
{{disable-model-invocation: true}}
---

# {{Skill Title}}

{{Brief statement of what the skill does and its goal. Define concrete
success artifacts — not just "writing code" but "a passing test suite
and a committed implementation."}}

## Inputs

- `$ARGUMENTS`: {{Description of expected arguments, or "(Optional) ..."}}

## Steps

### 1. {{Step Name}}

{{Specific, actionable instructions. Include commands when appropriate.}}

**Success criteria**: {{REQUIRED on every step. What proves this step
is done and we can move on.}}

Step annotations (key examples — consult skills-assist for complete conventions):

  • Success criteria is REQUIRED on every step
  • Execution: Direct (default — omit if direct), Delegate to [agent] (e.g., "Delegate to foundation:explorer"), or [human] (user does it)
  • Artifacts: Data this step produces that later steps need (PR number, commit SHA, file path). Only include if later steps depend on it.
  • Human checkpoint: Pause and ask before irreversible actions (merging, deploying, sending messages, destructive operations)
  • Rules: Hard constraints. User corrections during the original session are especially useful here.

Structural conventions:

  • Steps that can run concurrently use sub-numbers: 3a, 3b
  • Steps requiring the user to act get [human] in the title
  • Keep simple skills simple — a 2-step skill doesn't need annotations on every step

Frontmatter rules (key examples — consult skills-assist for complete reference):

  • description must carry all routing weight — the trigger, the "USE WHEN..." deciding factor, and the "DO NOT USE WHEN... — use <name>" boundary belong here, since this is what the visibility hook shows to the model. State a trigger as a decision rule, never as a worked <example> dialogue
  • allowed-tools should be the minimum set needed
  • context: fork only for self-contained skills that don't need user steering
  • If forked, usually pair with disable-model-invocation: true
  • model_role should match the skill's primary cognitive task
  • user-invocable: true registers the skill as a /name slash command
  • Names must be kebab-case, max 64 characters

Before you hand the skill back, measure the description you just wrote:

python3 -c "import yaml,sys; d=yaml.safe_load(open('SKILL.md').read().split('---')[1])['description']; print(len(d), 'chars'); sys.exit(len(d) > 400)"

If it exceeds 400, cut advocacy first (prose that sells the skill rather than routing to it), then compress the capability sentence. Never cut a trigger or a DO-NOT-USE boundary to make the number — a description that got shorter by dropping a routing fact is a mis-routing waiting to happen, and it surfaces later as "it didn't use the right thing" with nothing pointing back here.

Success criteria: The complete SKILL.md content has been drafted, and its description measures at or below 400 characters.

5. Test the Skill

Before committing, verify the skill works:

  1. Save the skill to .amplifier/skills/<name>/SKILL.md (immediately discoverable, no config changes needed).

  2. In the same session, call load_skill("<name>") and verify:

    • The skill loads without errors
    • The description is clear enough for routing
    • The body instructions are actionable
  3. Optionally, spawn a test session via delegate(agent="self", context_depth="none") that loads the skill and follows its instructions against a synthesized test input.

  4. If issues are found, fix and re-test.

Success criteria: The skill loads correctly and its instructions are actionable.

6. Review and Save

Before writing the file, present the complete SKILL.md content in a yaml code block so the user can review it with proper syntax highlighting.

Ask the user: "Does this look good to save?"

After confirmation:

  1. Create the skill directory:

    mkdir -p <chosen-path>/<skill-name>
    
  2. Write the SKILL.md file to <chosen-path>/<skill-name>/SKILL.md

  3. If the skill references companion files (templates, examples, scripts), create those as well.

  4. Tell the user:

    • Where the skill was saved
    • How to invoke it: /skill-name [arguments]
    • That they can edit the SKILL.md directly to refine it
    • If saved to a project or personal directory, the skill will be auto-discovered on next session start

Success criteria: The file is written to disk and the user knows how to use it.

来自 microsoft 的更多技能

oss-growth
microsoft
OSS增长黑客角色
agent-framework-azure-ai-py
microsoft
使用Microsoft Agent Framework Python SDK(agent-framework-azure-ai)构建Azure AI Foundry代理。在创建使用AzureAIAgentsProvider的持久化代理、使用托管工具(代码解释器、文件搜索、网络搜索)、集成MCP服务器、管理对话线程或实现流式响应时使用。涵盖函数工具、结构化输出和多工具代理。
development
airunway-aks-setup
microsoft
在AKS上设置AI Runway——从裸集群到运行模型。涵盖集群验证、控制器安装、GPU评估、提供商设置和首次部署。适用场景:“设置AI Runway”、“接入AKS集群”、“安装AI Runway”、“airunway设置”、“将模型部署到AKS”、“在AKS上进行GPU推理”、“在AKS上配置KAITO”、“在AKS上运行LLM”、“在AKS上使用vLLM”、“在AKS上设置模型服务”、“AI Runway控制器”。
devops
appinsights-instrumentation
microsoft
使用Azure Application Insights对Web应用进行插桩的指南。提供遥测模式、SDK设置和配置参考。适用场景:如何对应用进行插桩、App Insights SDK、遥测模式、什么是App Insights、Application Insights指南、插桩示例、APM最佳实践。
devops
applicationinsights-web-ts
microsoft
使用Application Insights JavaScript SDK(@microsoft/applicationinsights-web)为浏览器/Web应用添加检测。用于真实用户监控(RUM)——页面视图、点击、AJAX/fetch依赖项、异常、自定义事件,以及与后端OpenTelemetry追踪关联的浏览器端GenAI代理追踪。涵盖SDK加载器脚本和npm设置、框架扩展(React、React Native、Angular)、点击分析、遥测初始化器,以及从浏览器发出的代理/工具/模型跨度所遵循的OTel GenAI语义约定。
devops
azure-ai-anomalydetector-java
microsoft
使用适用于 Java 的 Azure AI 异常检测器 SDK 构建异常检测应用程序。在实现单变量/多变量异常检测、时间序列分析或 AI 驱动的监控时使用。
development
azure-ai-language-conversations-py
microsoft
使用azure-ai-language-conversations Python SDK实现对话语言理解(CLU)。当使用ConversationAnalysisClient分析对话意图和实体、构建NLP功能或将语言理解集成到应用程序中时使用。
development
azure-ai-ml-py
microsoft
Azure Machine Learning SDK v2 for Python。用于机器学习工作区、作业、模型、数据集、计算资源和管道。 触发词:“azure-ai-ml”、“MLClient”、“工作区”、“模型注册表”、“训练作业”、“数据集”。
development