evolve

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Evolve

Four modes: Session (run at session end), Personalize (discover and integrate external skills), Registry (curate

skillpacks/skill_dictionary.yaml

and presets), Create (build a new skill from scratch). Skill creation and significant rewrites always go through

skill-creator

— never write to SKILL.md directly.

skill-creator

is unavailable in the current environment, stop and tell the user. Do not silently simulate its workflow.

四种模式：Session（会话结束时运行）、Personalize（发现并集成外部技能）、Registry（管理

skillpacks/skill_dictionary.yaml

和预设）、Create（从零开始构建新技能）。技能创建和重大重写必须通过

skill-creator

完成——绝不直接写入SKILL.md。

如果当前环境中无法使用

skill-creator

，请停止操作并告知用户。请勿静默模拟其工作流程。

Mode 1 — Session

模式1 — 会话模式

Inputs

输入

Read

RESEARCH.md

History section and the git log (

git log --since="24 hours ago" --oneline

读取RESEARCH.md的History章节以及git日志（
```
git log --since="24 hours ago" --oneline
```
）。

Identify Candidate Lessons

识别候选经验

Look for: things that failed and why; unexpected experiment results; debugging insights that took significant effort; workflow improvements; decisions that were revisited or reversed.

寻找以下内容：失败的事项及原因；意外的实验结果；花费大量精力得到的调试见解；工作流程改进；被重新审视或推翻的决策。

Generalize Filter

普适性筛选

For each candidate: "Would this help in a different project on a different topic?"
- No → discard. Project-specific learnings stay in
```
RESEARCH.md
```
  .
- Yes → write a lesson file.

针对每个候选经验，思考：“这对不同主题的其他项目有帮助吗？”
- 否 → 舍弃。项目专属的学习内容保留在RESEARCH.md中。
- 是 → 撰写经验文件。

Write Lesson Files

撰写经验文件

Create
```
lessons/YYYYMMDD-<slug>.md
```
using
```
LESSON.md.template
```
. Fill in: Trigger, Lesson, Evidence, Applies to, Confidence (LOW/MED/HIGH), Status (PROPOSED).
If the lesson implies a skill change, classify it:
- Minor tweak (wording, one extra step, edge case) → write a unified diff under
```
## Proposed diff
```
  and a matching
```
lessons/YYYYMMDD-<slug>.diff
```
  .
- New skill or significant rewrite → hand off to
```
skill-creator
```
  : describe the intent, provide the lesson as context, and follow its draft→test→iterate loop. The resulting SKILL.md lands in
```
lessons/
```
  as a proposal — never installed directly.

使用LESSON.md.template创建
```
lessons/YYYYMMDD-<slug>.md
```
。填写以下内容：触发条件、经验内容、证据、适用场景、置信度（LOW/MED/HIGH）、状态（PROPOSED）。
若经验意味着需要变更技能，进行分类：
- 小幅调整（措辞修改、新增一个步骤、边缘场景处理）→ 在
```
## Proposed diff
```
  下编写统一差异文件，并生成对应的
```
lessons/YYYYMMDD-<slug>.diff
```
  文件。
- 新增技能或重大重写 → 移交至
```
skill-creator
```
  ：描述意图，提供经验作为上下文，并遵循其「草稿→测试→迭代」流程。生成的SKILL.md将作为提案存入
```
lessons/
```
  ——绝不直接安装。

Session Summary

会话总结

Print:

Session summary:
- TODOs completed: N
- Key metrics: <any that moved>
- Lessons extracted: N
- Skill diffs proposed: N

Tell the user: "Review proposed diffs in
lessons/
. Merge with
git apply
or edit directly."

打印：

Session summary:
- TODOs completed: N
- Key metrics: <any that moved>
- Lessons extracted: N
- Skill diffs proposed: N

告知用户：“请查看
lessons/
中的提案差异。可使用
git apply
合并或直接编辑。”

Mode 2 — Personalize

模式2 — 个性化模式

Activate when the user says "personalize", "integrate this skill", "find a skill for", or provides an external SKILL.md path or URL.

当用户说出“personalize”、“integrate this skill”、“find a skill for”，或提供外部SKILL.md的路径或URL时激活。

Step 0 — Discover (if no target provided)

步骤0 — 发现（未提供目标时）

If the user gives a description rather than a path/URL, search the registry:
bash
```
npx skills find "<query>"
```
Present the top results (name, description, install command) and ask the user which to integrate. Then continue with the chosen skill(s).

如果用户提供的是描述而非路径/URL，搜索注册表：
bash
```
npx skills find "<query>"
```
展示排名靠前的结果（名称、描述、安装命令），并询问用户要集成哪一个。随后继续处理所选技能。

Step 1 — Read external skill(s)

步骤1 — 读取外部技能

Read the target: a single SKILL.md, a skills directory, or a skill pack repo. Extract:
- Name, description, trigger phrases
- Tools, MCP servers, and external dependencies it references
- Behavioral patterns (workflows, delegation, output files)

读取目标内容：单个SKILL.md、技能目录或技能包仓库。提取以下信息：
- 名称、描述、触发短语
- 其引用的工具、MCP服务器和外部依赖
- 行为模式（工作流程、任务委托、输出文件）

Step 2 — Inventory installed skills

步骤2 — 盘点已安装技能

Read all SKILL.md files in
```
~/.claude/skills/
```
(or the project
```
.claude/skills/
```
). Build a map of:
- Names and descriptions
- Tools and MCP servers already in use
- Overlapping trigger phrases

读取
```
~/.claude/skills/
```
（或项目的
```
.claude/skills/
```
）中的所有SKILL.md文件。构建以下映射：
- 名称和描述
- 已在使用的工具和MCP服务器
- 重叠的触发短语

Step 3 — Analyze

步骤3 — 分析

For each external skill, assess three dimensions:

Compatibility — will it conflict?
- Name collision with an installed skill?
- References a tool/MCP not configured (e.g.,
```
alpha login
```
  not done)?
- Contradicts a behavioral rule in an existing skill?
Scope overlap — does it duplicate something?
- Substantially the same as an installed skill → flag as overlap, note the delta
- Partially overlapping → identify which parts are additive
New capability — what does it genuinely add?
- List capabilities not covered by any installed skill

针对每个外部技能，从三个维度进行评估：

兼容性 — 是否会产生冲突？
- 是否与已安装技能存在名称冲突？
- 是否引用了未配置的工具/MCP（例如未执行
```
alpha login
```
  ）？
- 是否与现有技能的行为规则相矛盾？
范围重叠 — 是否存在重复内容？
- 与已安装技能基本相同 → 标记为重叠，记录差异点
- 部分重叠 → 识别新增的内容部分
新能力 — 真正新增了什么？
- 列出未被任何已安装技能覆盖的能力

Step 4 — Interview the user

步骤4 — 与用户沟通

Ask targeted questions based on the analysis. Ask all at once, not one-by-one:
For each overlap: > "
```
<external>
```
overlaps with your
```
<installed>
```
. It differs in:
```
<delta>
```
. Replace, merge, or skip?"
For each missing dependency: > "
```
<external>
```
requires
```
<tool/MCP>
```
. Do you have it? If not, skip those parts?"
For each behavioral conflict: > "
```
<external>
```
does
```
<X>
```
but your
```
<installed>
```
does
```
<Y>
```
. Which do you prefer?"
Open: > "Anything from this skill you explicitly don't want?"

根据分析结果提出针对性问题。一次性全部提出，而非逐个询问：
针对每个重叠项：> "
```
<external>
```
与您已安装的
```
<installed>
```
存在重叠。差异点为：
```
<delta>
```
。是否替换、合并或跳过？"
针对每个缺失依赖：> "
```
<external>
```
需要
```
<tool/MCP>
```
。您是否已配置？若未配置，是否跳过相关部分？"
针对每个行为冲突：> "
```
<external>
```
会执行
```
<X>
```
，但您已安装的
```
<installed>
```
会执行
```
<Y>
```
。您偏好哪一种？"
开放式问题：> "您明确不想要该技能中的哪些内容？"

Step 5 — Create or improve via skill-creator

步骤5 — 通过skill-creator创建或改进

Based on user answers, determine the scope:
- Additive merge (minor additions to an existing skill) → write a unified diff to
```
lessons/YYYYMMDD-personalize-<slug>.diff
```
  with a companion
```
.md
```
  .
- New skill or substantial rewrite → hand off to
```
skill-creator
```
  : describe the desired skill, provide the source material and user's confirmed preferences as context, and follow its draft→test→iterate loop. Output lands in
```
lessons/
```
  as a proposal.
Tell the user: "Ready in
lessons/
. Review and apply with
git apply
when satisfied."

根据用户的回答确定范围：
- 增量合并（对现有技能进行小幅补充）→ 将统一差异写入
```
lessons/YYYYMMDD-personalize-<slug>.diff
```
  ，并生成对应的
```
.md
```
  文件。
- 新增技能或重大重写 → 移交至
```
skill-creator
```
  ：描述所需技能，提供源材料和用户确认的偏好作为上下文，并遵循其「草稿→测试→迭代」流程。输出结果作为提案存入
```
lessons/
```
  。
告知用户：“已准备就绪，存于
lessons/
中。满意后可使用
git apply
应用变更。”

Mode 3 — Registry

模式3 — 注册表模式

Activate when the user asks to refresh the tracked skill-pack catalog, assess a new pack, compare packs, or update

skillpacks/skill_dictionary.yaml

/ preset definitions.

当用户要求刷新跟踪的技能包目录、评估新技能包、比较技能包，或更新

skillpacks/skill_dictionary.yaml

/ 预设定义时激活。

Registry inputs

注册表输入

Read:

```
skillpacks/skill_dictionary.yaml
```
```
skillpacks/presets/*.yaml
```
```
README.md
```
/
```
README_CN.md
```
pack comparison sections when relevant

If the user points at a new external repo, read its README and any obvious skill inventory files first.

读取：

```
skillpacks/skill_dictionary.yaml
```
```
skillpacks/presets/*.yaml
```
相关的README.md / README_CN.md中的技能包比较章节

如果用户指向新的外部仓库，先读取其README和任何明显的技能清单文件。

Responsibilities

职责

Maintain
```
skillpacks/skill_dictionary.yaml
```
as the shared curated registry.
For each pack or skill being assessed, judge:
- workflow fit
- gap coverage
- overlap cost
- dependency burden
- maintenance health
- composability
- supervision sensitivity
- autonomy bias

Normalize outputs into concise labels when possible:

necessity:

default | recommended | optional | niche | avoid

compatibility:
```
high | medium | low
```

role:

base | donor | overlay | niche | experimental

Recommend subsets, not bulk imports, unless the user clearly asks for a full-pack strategy.
Update presets when the registry change meaningfully affects default recommendations.
Keep rationales concise and decision-oriented. The registry is for recommendation and comparison, not exhaustive catalog dumps.

将
```
skillpacks/skill_dictionary.yaml
```
维护为共享的精选注册表。
针对每个被评估的技能包或技能，判断：
- 工作流程适配性
- 缺口覆盖度
- 重叠成本
- 依赖负担
- 维护健康度
- 可组合性
- 监督敏感性
- 自主性偏差

尽可能将输出标准化为简洁标签：

必要性：

default | recommended | optional | niche | avoid

兼容性：
```
high | medium | low
```

角色：

base | donor | overlay | niche | experimental

推荐子集，而非批量导入，除非用户明确要求全技能包策略。
当注册表变更对默认推荐产生重大影响时，更新预设。
保持理由简洁且以决策为导向。注册表用于推荐和比较，而非详尽的目录转储。

Safety boundary

安全边界

Registry curation may edit
```
skillpacks/skill_dictionary.yaml
```
and
```
skillpacks/presets/*.yaml
```
, but changes to existing
```
SKILL.md
```
behavior still require either:
- a minor diff proposal in
```
lessons/
```
  , or
- a
```
skill-creator
```
  handoff for substantial rewrites.
If a registry update would imply changing default project behavior materially, ask before proceeding.

注册表管理可编辑
```
skillpacks/skill_dictionary.yaml
```
和
```
skillpacks/presets/*.yaml
```
，但要变更现有SKILL.md的行为仍需：
- 在
```
lessons/
```
  中提交小幅差异提案，或
- 对于重大重写，移交至
```
skill-creator
```
  处理。
若注册表更新会实质性改变默认项目行为，需先询问用户再进行操作。

Mode 4 — Create

模式4 — 创建模式

Activate when the user says "create a skill for X" or "build a new skill".

Capture intent: what should the skill do, when should it trigger, what's the expected output?
Check
```
npx skills find "<intent>"
```
— if a close match exists, suggest integrating it instead (Mode 2).
If creating from scratch, hand off to
```
skill-creator
```
with the captured intent. Follow its draft→test→iterate loop fully. Output lands in
```
lessons/
```
as a proposal.

当用户说出“create a skill for X”或“build a new skill”时激活。

捕捉意图：该技能应实现什么功能、何时触发、预期输出是什么？
执行
```
npx skills find "<intent>"
```
——若存在相近匹配项，建议改为集成该技能（模式2）。
若需从零开始创建，将捕捉到的意图移交至
```
skill-creator
```
。完整遵循其「草稿→测试→迭代」流程。输出结果作为提案存入
```
lessons/
```
。

Constraints

约束条件

Never write directly to any
```
SKILL.md
```
file. Always propose. The human merges.
Create
```
lessons/
```
if it does not exist.
```
skillpacks/skill_dictionary.yaml
```
and
```
skillpacks/presets/*.yaml
```
are curated repo metadata and may be updated directly in Registry mode when the user asks.
All skill creation and significant rewrites go through
```
skill-creator
```
. Raw diffs are only for minor, targeted changes.
Err toward asking rather than assuming. One unanswered question beats a wrong assumption baked into a diff.

绝不直接写入任何SKILL.md文件。始终以提案形式提交，由人工合并。
若
```
lessons/
```
目录不存在则创建。
```
skillpacks/skill_dictionary.yaml
```
和
```
skillpacks/presets/*.yaml
```
是精选的仓库元数据，当用户要求时，可在注册表模式下直接更新。
所有技能创建和重大重写均需通过
```
skill-creator
```
完成。仅针对小幅、定向变更使用原始差异文件。
优先询问而非假设。一个未解答的问题好过一个错误假设被写入差异文件。

Example

示例

Session mode: Experiment failed due to unset random seed. Minor →

lessons/20260329-seed-reproducibility.diff

. New workflow implied → hand off to

skill-creator

to draft a

reproducibility-check

skill.

Personalize mode: User says "find me a skill for LaTeX compilation" →

npx skills find "latex"

returns 3 results → user picks one → reads it, finds no conflicts → substantial enough to need skill-creator → draft→test→iterate →

lessons/20260401-personalize-latex.diff

Create mode: User says "create a skill for Slack notifications" →

npx skills find "slack"

finds a close match → suggests integrating it → user confirms → Mode 2 kicks in.

会话模式：实验因未设置随机种子而失败。属于小幅调整 → 生成

lessons/20260329-seed-reproducibility.diff

。若涉及新工作流程 → 移交至skill-creator以起草

reproducibility-check

技能。

个性化模式：用户说“find me a skill for LaTeX compilation” →

npx skills find "latex"

返回3个结果 → 用户选择其中一个 → 读取该技能，未发现冲突 → 内容足够复杂，需skill-creator处理 → 草稿→测试→迭代 → 生成

lessons/20260401-personalize-latex.diff

。

创建模式：用户说“create a skill for Slack notifications” →

npx skills find "slack"

找到相近匹配项 → 建议集成该技能 → 用户确认 → 启动模式2。