eve-skill-distillation

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Eve Skill Distillation

Eve技能提炼

Use this workflow to turn repeated patterns into reusable skills.

使用此工作流将重复模式转化为可复用的技能。

When to distill

何时进行提炼

A workflow has repeated across two or more jobs.
Knowledge would benefit other agents working on the same project.
A failure mode keeps recurring and the fix should be encoded.

某一工作流程在两个或多个任务中重复出现。
相关知识对同一项目中的其他Agent有帮助。
某一故障模式反复出现，其修复方法应被固化为技能。

Workflow: Orchestrate, Don't Serialize

工作流：编排而非串行

When distilling involves multiple skills (creating several, updating a batch, or a mix), use an orchestrator pattern rather than doing everything sequentially. This protects your context budget and parallelizes independent work.

当提炼涉及多个技能（创建多个技能、批量更新或两者结合）时，请使用编排器模式，而非按顺序执行所有操作。这能节省上下文开销，并并行处理独立任务。

Step 1: Capture patterns (orchestrator)

步骤1：捕获模式（编排器）

Identify all the patterns worth distilling from recent work. For each one:

Name the repeated steps, commands, or failure modes.
Decide: update an existing skill, or create a new one?
Choose the target repo/pack first - is this likely to be
- ```
./private-skills/
```
  or similar for repo-specific development workflows that are not going to be imported into other projects (private)
- ```
./private-eve-dev-skills/
```
  when working in the eve-horizon repo and the skills relate to internal development of the platform itself (private)
- ```
../eve-skillpacks/eve-se/
```
  for Eve platform work (deploy, auth, manifests, pipelines)
- ```
../eve-skillpacks/eve-work/
```
  for general knowledge work (docs, orchestration, distillation)
- ```
../eve-skillpacks/eve-design/
```
  for architecture and design thinking

从近期工作中找出所有值得提炼的模式。针对每个模式：

为重复的步骤、命令或故障模式命名。
决定：是更新现有技能，还是创建新技能？
首先选择目标仓库/技能包——该技能属于以下哪一类？
- ```
./private-skills/
```
  或类似目录：用于特定仓库的开发工作流，不会导入到其他项目中（私有）
- ```
./private-eve-dev-skills/
```
  ：当在eve-horizon仓库中工作，且技能与平台自身的内部开发相关时使用（私有）
- ```
../eve-skillpacks/eve-se/
```
  ：用于Eve平台相关工作（部署、认证、清单、流水线）
- ```
../eve-skillpacks/eve-work/
```
  ：用于通用知识工作（文档、编排、提炼）
- ```
../eve-skillpacks/eve-design/
```
  ：用于架构与设计思维

Step 2: Plan work items (orchestrator)

步骤2：规划工作项（编排器）

Create a tracked work item for each skill to create or update. Each work item description must be self-contained — a worker with no prior context should be able to execute it.

Include in each work item:

The target file path (e.g.,
```
eve-se/eve-auth-and-secrets/SKILL.md
```
)
Whether this is a create or update
The pattern being captured: what the skill should teach
Any source material the worker should read (existing skills, reference docs, conversation history)
The authoring rules (see "Skill Authoring Rules" below)

If there are housekeeping updates (README, ARCHITECTURE.md), add those as a final work item blocked until all skill work items complete.

为每个需要创建或更新的技能创建一个可追踪的工作项。每个工作项的描述必须独立完整——即使是没有前置上下文的Worker也能执行。

每个工作项需包含：

目标文件路径（例如：
```
eve-se/eve-auth-and-secrets/SKILL.md
```
）
操作类型：创建或更新
要捕获的模式：该技能应传授的内容
Worker需要阅读的所有源材料（现有技能、参考文档、对话历史）
创作规则（见下文“技能创作规则”）

如果有维护更新（如README、ARCHITECTURE.md），将其添加为最后一个工作项，需等待所有技能工作项完成后再执行。

Step 3: Dispatch workers (parallel)

步骤3：分派Worker（并行）

Spawn one worker per skill work item. Launch them all at once.

Each worker prompt must be self-contained. The worker has no access to the orchestrator's conversation. Include:

The target file path to create or edit
The pattern to capture, described in enough detail to write from
The authoring rules below
Existing file content to preserve (for updates)

为每个技能工作项分配一个Worker，同时启动所有Worker。

每个Worker的提示必须独立完整。Worker无法访问编排器的对话内容，提示需包含：

要创建或编辑的目标文件路径
要捕获的模式，描述需足够详细以便撰写技能内容
下文的创作规则
需要保留的现有文件内容（针对更新操作）

Step 4: Collect and finalize (orchestrator)

步骤4：收集并定稿（编排器）

Wait for all workers to complete. Then:

Update pack README and
```
ARCHITECTURE.md
```
listings if skills were added or removed.
Verify each new/updated skill follows the authoring rules.

等待所有Worker完成任务，然后：

如果有技能添加或移除，更新技能包的README和
```
ARCHITECTURE.md
```
中的列表。
验证每个新创建/更新的技能是否符合创作规则。

Single-skill shortcut

单技能快捷方式

If there's only one skill to distill, skip the orchestrator pattern and do it directly. The overhead of dispatching a single worker isn't worth it.

如果仅需提炼一个技能，可跳过编排器模式直接操作。为单个Worker分派任务的开销得不偿失。

Skill Authoring Rules

技能创作规则

Every worker (or direct author) must follow these:

Frontmatter: YAML with
```
name
```
and
```
description
```
only.
Voice: Imperative form throughout. ("Run the command", not "You should run the command".)
Conciseness: Keep SKILL.md under 5,000 words. Move long details into
```
references/
```
.
Teach thinking, not just steps: Skills should help agents understand why, not just what. Include the reasoning behind workflows so agents can adapt when conditions change.
Structure for skimming: Use headers, short paragraphs, and code blocks. Agents scan before they read.
Agent-agnostic language: Describe what to do, not which tool to call. Say "edit the file" not "use the Edit tool". Say "spawn a background worker" not "launch a Task sub-agent".

所有Worker（或直接创作者）必须遵循以下规则：

前置内容：仅包含
```
name
```
和
```
description
```
的YAML格式内容。
语气：全程使用祈使语气。（例如：“运行命令”，而非“你应该运行命令”。）
简洁性：SKILL.md文件字数控制在5000字以内。将冗长细节移至
```
references/
```
目录中。
传授思维，而非仅步骤：技能应帮助Agent理解为什么，而非仅做什么。需包含工作流背后的推理逻辑，以便Agent在条件变化时能够调整操作。
便于浏览的结构：使用标题、短段落和代码块。Agent会先浏览再阅读。
与Agent无关的语言：描述要做什么，而非调用哪个工具。例如说“编辑文件”而非“使用编辑工具”；说“启动后台Worker”而非“启动Task子Agent”。

Recursive distillation

递归提炼

Repeat this loop after each significant job.
Merge overlapping skills instead of duplicating them.
Keep skills current as platform behavior evolves.
When a skill's instructions no longer match reality, update or retire it.

在每个重要任务完成后重复此循环。
合并重叠的技能，避免重复。
随着平台行为的演变，保持技能的时效性。
当技能的说明与实际情况不符时，更新或淘汰该技能。