Back to Details

start

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese
<EXTREMELY-IMPORTANT> This skill is a lightweight router, not a universal constitution.
Non-negotiable rules:
  1. If a more specific skill matches, invoke it before giving a substantive task response.
  2. Keep intake lightweight. Do not enumerate all skills, agents, and plugins unless the user explicitly asks for a capability inventory.
  3. Gather only the minimum context needed to avoid blind edits or duplicated work.
  4. Delegate to a specialized agent only when domain expertise or isolation materially improves the result.
  5. Use planning and task tracking only when the work is genuinely multi-step, risky, or ambiguous.
</EXTREMELY-IMPORTANT>
<EXTREMELY-IMPORTANT> 该skill是一个轻量级路由，并非通用规则体系。
不可协商的规则：
  1. 如果有更匹配的特定skill，应在给出实质性任务响应前调用它。
  2. 保持接收流程轻量化。除非用户明确要求能力清单，否则不要枚举所有skill、Agent和插件。
  3. 仅收集避免盲目编辑或重复工作所需的最少上下文。
  4. 只有当领域专业知识或隔离能显著提升结果时，才委派给专业Agent。
  5. 仅当工作确实是多步骤、有风险或模糊不清时，才使用规划和任务跟踪。
</EXTREMELY-IMPORTANT>

Start

启动流程

Goal

目标

Choose the next best execution path at the beginning of a task with minimal prompt overhead.
This skill should answer four questions quickly:
  1. Is a more specific skill the real workflow?
  2. Is specialized agent delegation warranted?
  3. What is the minimum context needed before acting?
  4. Does the task need planning, or can it proceed directly?
以最小的提示开销，在任务开始时选择最佳的下一步执行路径。
该skill应快速回答四个问题：
  1. 更特定的skill是否是真正的工作流？
  2. 是否需要委派给专业Agent？
  3. 执行前所需的最少上下文是什么？
  4. 任务是否需要规划，还是可以直接执行？

Step 1: Check for a more specific skill

步骤1：检查是否有更特定的skill

Match the user request against domain and workflow skills first.
  • If a skill clearly applies, invoke it before explaining the approach.
  • If multiple skills may apply, choose the smallest set that covers the task.
  • If you are unsure whether a skill matches, read 
    references/skill_discovery_patterns.md
    .
  • Do not perform full filesystem skill discovery on every task. Claude already exposes available skills and treats matching skill invocation as a first-class workflow.
Success criteria: You either invoked the correct skill, or you can state concretely why no more specific skill applies.
首先将用户请求与领域和工作流skill进行匹配。
  • 如果某个skill明显适用，应在解释方法前调用它。
  • 如果多个skill可能适用，选择覆盖任务的最小集合。
  • 如果不确定某个skill是否匹配，请阅读
    references/skill_discovery_patterns.md
    。
  • 不要在每个任务中都执行完整的文件系统skill发现。Claude已暴露可用skill，并将匹配的skill调用作为一等工作流处理。
成功标准：要么调用了正确的skill，要么能具体说明为何没有更特定的skill适用。

Step 2: Decide whether to delegate to a specialized agent

步骤2：决定是否委派给专业Agent

Use an agent only when specialization or isolation changes the outcome materially.
Delegate when:
  • the task is domain-heavy implementation, debugging, or review work
  • a specialized agent has clearly better heuristics for the stack
  • the work is large enough that ownership boundaries help
Do not delegate when:
  • the work is a trivial read, search, or small direct edit
  • the task is only capability discovery
  • the next step depends on immediate local inspection
If agent choice is unclear, read 
references/agent_matching_logic.md
.
Success criteria: You chose either direct execution or a specific agent for a concrete reason.
仅当专业化或隔离能显著改变结果时才使用Agent。
在以下情况进行委派：
  • 任务是领域密集型的实现、调试或评审工作
  • 专业Agent针对该技术栈有明显更优的启发式策略
  • 工作规模足够大，所有权边界有助于推进
在以下情况不进行委派：
  • 工作是简单的读取、搜索或小型直接编辑
  • 任务仅为能力发现
  • 下一步依赖即时本地检查
如果Agent选择不明确，请阅读
references/agent_matching_logic.md
。
成功标准：出于具体原因选择了直接执行或特定Agent。

Step 3: Gather minimum viable context

步骤3：收集最小可行上下文

Inspect only the code and files needed to avoid blind action.
  • Identify the likely files, modules, or subsystems involved.
  • Read surrounding code before editing.
  • Search for existing implementations or patterns before building from scratch.
  • Escalate to broader exploration only if the request is ambiguous or the code surface is unclear.
Use 
references/discovery.md
only when:
  • the user explicitly asks what skills, agents, or plugins are available
  • the environment appears out of sync with the file system
  • you need a capability inventory as the actual task output
Success criteria: You have enough context to act without forcing broad discovery or redundant reading.
仅检查避免盲目操作所需的代码和文件。
  • 确定可能涉及的文件、模块或子系统。
  • 编辑前阅读相关代码。
  • 从零开始构建前，搜索现有实现或模式。
  • 仅当请求模糊或代码范围不明确时，才扩大探索范围。
仅在以下情况使用
references/discovery.md
：
  • 用户明确询问可用的skill、Agent或插件
  • 环境似乎与文件系统不同步
  • 需要将能力清单作为实际任务输出
成功标准：拥有足够的上下文来执行操作，无需强制进行大范围发现或冗余阅读。

Step 4: Choose planning depth

步骤4：选择规划深度

Pick the smallest planning mechanism that keeps the work safe.
  • Direct execution: trivial, well-bounded requests
  • Light task tracking: several steps or files, but straightforward work
  • Full DAG planning: ambiguous, risky, architectural, or parallelizable work
If planning is the real task, invoke 
plan-to-task-list-with-dag
instead of turning 
start
into a giant planner.
Success criteria: The task has an execution shape that matches its real complexity.
选择能保证工作安全的最小规划机制。
  • 直接执行：简单、边界清晰的请求
  • 轻量任务跟踪：涉及多个步骤或文件，但工作流程明确
  • 完整DAG规划：模糊、有风险、架构性或可并行的工作
如果规划是实际任务，请调用
plan-to-task-list-with-dag
，而不是将
start
变成一个大型规划器。
成功标准：任务的执行模式与其实际复杂度匹配。

Step 5: Communicate the next move

步骤5：传达下一步行动

Give a concise first update that states:
  • what you think the user wants
  • what you are checking first
  • whether you are invoking another skill, delegating to an agent, or proceeding directly
If delegating, pass:
  • clear scope
  • relevant context only
  • success criteria
If executing directly, begin with the highest-value context step immediately after the update.
Success criteria: The user can see the execution path without reading a long preamble.
给出简洁的首次更新，说明：
  • 你对用户需求的理解
  • 你首先要检查的内容
  • 是否要调用另一个skill、委派给Agent，还是直接执行
如果进行委派，请传递：
  • 明确的范围
  • 仅相关上下文
  • 成功标准
如果直接执行，请在更新后立即开始最高价值的上下文步骤。
成功标准：用户无需阅读冗长的前言即可了解执行路径。

Guardrails

防护规则

  • Do not announce a skill without invoking it.
  • Do not use 
    start
    as a substitute for a real domain skill.
  • Do not force full capability discovery on normal task intake.
  • Do not sync 
    CLAUDE.md
    as part of routine startup.
  • Do not create task-tracking noise for trivial work.
  • Do not delegate simple search or file-inspection work that should stay local.
  • 不要只宣布skill而不调用它。
  • 不要用
    start
    替代真正的领域skill。
  • 不要在常规任务接收时强制进行完整的能力发现。
  • 不要将同步
    CLAUDE.md
    作为常规启动流程的一部分。
  • 不要为简单工作创建任务跟踪噪音。
  • 不要委派应在本地完成的简单搜索或文件检查工作。

When To Load References

何时加载参考文档

  • references/skill_discovery_patterns.md
    Use when unsure whether a skill applies, or when several skills overlap.
  • references/agent_matching_logic.md
    Use when unsure which specialized agent fits the task.
  • references/discovery.md
    Use only for explicit capability inventory or environment-drift checks.
  • references/sync-claude-md.md
    Use only when the user explicitly wants the capability section of 
    CLAUDE.md
    refreshed and the session permits file edits.
  • references/skill_discovery_patterns.md
    当不确定某个skill是否适用，或多个skill重叠时使用。
  • references/agent_matching_logic.md
    当不确定哪个专业Agent适合任务时使用。
  • references/discovery.md
    仅用于明确的能力清单或环境漂移检查。
  • references/sync-claude-md.md
    仅当用户明确要求刷新
    CLAUDE.md
    的能力部分且会话允许文件编辑时使用。

Output Contract

输出约定

On normal invocation, keep the first response short and concrete:
  1. state the task understanding
  2. state the next action
  3. name the invoked skill or delegated agent if one is being used
This skill succeeds when it routes cleanly and then gets out of the way.
正常调用时，首次响应应简短具体：
  1. 说明对任务的理解
  2. 说明下一步行动
  3. 若要调用skill或委派Agent，指明其名称
当该skill能清晰完成路由并让开执行路径时，即视为成功。