curate-skills

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Skill: Curate Skills

Skill：Skill治理

Purpose

目标

Govern the skill inventory by evaluating, scoring, tagging, and normalizing every Skill in the repository (including this one). Produce a single source of truth: machine-readable scores and status per skill, normalized human-facing README, overlap detection, and a repo-level summary. Agent-first; README for humans, agent.yaml for Agents.

通过对仓库中所有Skill（包括本Skill）进行评估、打分、打标签和标准化处理，对Skill资产进行管控。生成唯一可信数据源：机器可读的各Skill分数与状态、面向人类的标准化README、重叠检测结果，以及仓库级摘要。以Agent为核心；README面向人类，agent.yaml面向Agent。

Use Cases

适用场景

After adding or changing skills: Re-score and update status and docs so the inventory stays consistent.
Audit: Review all skills for lifecycle (validated / experimental / archive_candidate) and overlap.
Repo summary: Generate or refresh ASQM_AUDIT.md or a structured chat summary for the whole skills directory.
Self-evaluation: Run curation on the repo including this meta-skill so governance is itself a skill.

新增或修改Skill后：重新打分并更新状态与文档，确保Skill资产保持一致性。
审计：检查所有Skill的生命周期状态（已验证 / 实验性 / 待归档候选）与重叠情况。
仓库摘要：为整个Skill目录生成或更新ASQM_AUDIT.md，或结构化聊天摘要。
自我评估：对包含本元Skill在内的仓库执行治理，让管控本身也成为一项Skill。

Behavior

行为逻辑

Scan: List all skill directories under the given
```
skills_directory
```
(e.g.
```
skills/
```
).
Read: For each skill, read
```
agent.yaml
```
if present; otherwise read README or SKILL.md. Prefer agent.yaml when it exists.
Score: For each skill, assign four ASQM scores 0–5 strictly:
```
agent_native
```
,
```
cognitive
```
,
```
composability
```
,
```
stance
```
. Apply strict scoring (evidence-based, no inflation; see below). Compute Quality (linear):
```
asqm_quality = agent_native + cognitive + composability + stance
```
(0–20). Write scores and asqm_quality to
```
agent.yaml
```
.
Lifecycle: Apply dual-gate rules. Gate A (agent readiness): agent_native ≥ 4. Gate B (design integrity): stance ≥ 3. validated: Quality ≥ 17 AND Gate A AND Gate B. experimental: Quality ≥ 10. archive_candidate: otherwise.

Overlaps and position: For each skill, assign

overlaps_with

(list of overlapping skills in Git-repo form

owner/repo:skill-name

, same format for this repo and others; e.g.

nesnilnehc/ai-cortex:generate-standard-readme

softaworks/agent-toolkit:commit-work

) and

market_position

(

differentiated

commodity

experimental

Write: Per skill, write or update
```
agent.yaml
```
(scores, status, overlaps_with, market_position) and normalize
```
README.md
```
to standard sections (what it does, when to use, inputs/outputs, etc.).
Summary: Either write
```
ASQM_AUDIT.md
```
at repo level or print a structured summary in chat. ASQM_AUDIT.md must include a final recommendations section (e.g. “Recommendation”): actionable next steps (e.g. score adjustments, SKILL changes) or an explicit “no changes recommended” conclusion, so every audit ends with clear guidance. Required sections: lifecycle by status, scoring formula, dimension checklist, overlaps, ecosystem, findings, recommendations (final), and a short summary table.

Conceptual split

Scores (ASQM) measure intrinsic quality: how well the skill is designed for Agents, reasoning offloaded, composability, and stance.
Overlaps and market_position describe ecosystem position: how the skill relates to others in the inventory.

Scoring model: ASQM (linear quality + dual-gate)

Quality (linear):

asqm_quality = agent_native + cognitive + composability + stance

; each dimension 0–5, total 0–20.

Gate A (agent readiness): agent_native ≥ 4.
Gate B (design integrity): stance ≥ 3.
Lifecycle: validated ↔ Quality ≥ 17 AND Gate A AND Gate B; experimental ↔ Quality ≥ 10; archive_candidate ↔ otherwise. (Bar set so validated = clearly production-ready: 17/20 + both gates.)
Dimensions: agent_native — Agent consumption (contracts, machine-readable metadata). cognitive — Reasoning offloaded from user to Agent. composability — Ease of combining with other skills or pipelines. stance — Design stance (spec alignment, principles).

Strict scoring (required)

Evidence-based: Each score must be justified by the skill’s SKILL.md (e.g. presence of Appendix: Output contract, related_skills, Restrictions, Self-Check).
No inflation: agent_native = 5 only when the skill has an explicit, machine-parseable output contract (e.g. Appendix: Output contract or equivalent table/spec in SKILL.md). If output is described only in prose, agent_native ≤ 4.
Consistency: Apply the same criteria across all skills; do not relax for a single skill without justification.

Ecosystem position (per skill)

overlaps_with: List of overlapping skills in Git-repo form
```
owner/repo:skill-name
```
. Use the same format for this repo and for other repos (no separate internal/external). Examples:
```
nesnilnehc/ai-cortex:refine-skill-design
```
,
```
softaworks/agent-toolkit:commit-work
```
. Empty
```
[]
```
if none.
market_position:
- ```
differentiated
```
  : Clear differentiator, minimal overlap, distinct value in the inventory.
- ```
commodity
```
  : Common capability, overlaps with many skills, standard pattern.
- ```
experimental
```
  : Early-stage, niche, or unclear positioning in the ecosystem.

Interaction: Before overwriting many skill files or writing ASQM_AUDIT.md, confirm with the user unless they have explicitly requested a full run (e.g. “curate all skills” or “run curate-skills”).

扫描：列出指定
```
skills_directory
```
（如
```
skills/
```
）下的所有Skill目录。
读取：针对每个Skill，若存在
```
agent.yaml
```
则优先读取；否则读取README或SKILL.md。优先使用agent.yaml内容。
打分：为每个Skill严格分配0-5分的四项ASQM评分：
```
agent_native
```
、
```
cognitive
```
、
```
composability
```
、
```
stance
```
。采用严格打分规则（基于证据，不得虚高；详见下文）。计算质量分（线性求和）：
```
asqm_quality = agent_native + cognitive + composability + stance
```
（0-20分）。将分数与asqm_quality写入
```
agent.yaml
```
。
生命周期判定：应用双门槛规则。门槛A（Agent就绪度）：agent_native ≥4。门槛B（设计完整性）：stance ≥3。已验证：质量分≥17 且满足门槛A与门槛B。实验性：质量分≥10。待归档候选：其余情况。

重叠与定位：为每个Skill分配

overlaps_with

（重叠Skill列表，采用Git仓库格式

owner/repo:skill-name

，本仓库与外部仓库格式一致；例如

nesnilnehc/ai-cortex:generate-standard-readme

、

softaworks/agent-toolkit:commit-work

）与

market_position

（

differentiated

commodity

experimental

）。

写入：为每个Skill编写或更新
```
agent.yaml
```
（包含分数、状态、overlaps_with、market_position），并将
```
README.md
```
标准化为固定章节（功能介绍、适用场景、输入输出等）。
摘要生成：在仓库根目录生成ASQM_AUDIT.md，或在聊天中输出结构化摘要。ASQM_AUDIT.md必须包含最终建议章节（如“Recommendation”）：可执行的后续步骤（如分数调整、Skill修改）或明确的“无需修改”结论，确保每次审计都给出清晰指引。必填章节：按状态分类的生命周期、打分公式、维度检查清单、重叠情况、生态系统、审计发现、最终建议，以及简短的汇总表格。

概念区分

ASQM评分：衡量Skill的内在质量：是否为Agent原生设计、认知能力、可组合性、设计立场。
重叠与市场定位：描述Skill在生态中的位置：与其他Skill的重叠关系、在资产库中的定位。

严格打分规则（强制执行）

基于证据：每项分数必须有Skill的SKILL.md内容支撑（如附录：输出契约、关联Skill、限制条件、自我检查项的存在）。
不得虚高：仅当Skill具备明确的、机器可解析的输出契约（如SKILL.md中的附录：输出契约或等效表格/规范）时，agent_native才可打5分。若仅用自然语言描述输出，agent_native最高为4分。
一致性：对所有Skill应用相同打分标准；不得无理由为单个Skill放宽规则。

生态定位（按单个Skill）

overlaps_with：重叠Skill列表，采用Git仓库格式
```
owner/repo:skill-name
```
。本仓库与外部仓库格式统一，无内外区分。示例：
```
nesnilnehc/ai-cortex:refine-skill-design
```
、
```
softaworks/agent-toolkit:commit-work
```
。无重叠则为
```
[]
```
。
market_position:
- ```
differentiated
```
  ：具备明显差异化，重叠少，在资产库中拥有独特价值。
- ```
commodity
```
  ：通用能力，与多个Skill重叠，属于标准模式。
- ```
experimental
```
  ：早期阶段、小众场景，或在生态中定位尚不清晰。

交互规则：除非用户明确要求全量执行（如“治理所有Skill”或“运行curate-skills”），否则在覆盖大量Skill文件或生成ASQM_AUDIT.md前，需先与用户确认。

Input & Output

输入与输出

Input

```
skills_directory
```
: Root path containing skill subdirectories (e.g.
```
skills/
```
).

Output

Per skill: updated
```
agent.yaml
```
(scores, status, overlaps_with, market_position); normalized
```
README.md
```
.
Repo-level:
```
ASQM_AUDIT.md
```
or structured summary in chat.
Overlap and market_position report: per-skill overlaps_with (owner/repo:skill-name), market_position.

输入

```
skills_directory
```
：包含Skill子目录的根路径（如
```
skills/
```
）。

输出

单Skill层面：更新后的
```
agent.yaml
```
（包含分数、状态、overlaps_with、market_position）；标准化后的
```
README.md
```
。
仓库层面：
```
ASQM_AUDIT.md
```
或聊天中的结构化摘要。
重叠与定位报告：每个Skill的overlaps_with（owner/repo:skill-name）与market_position。

Restrictions

限制条件

Do not change spec/skill.md or manifest.json from within this skill; metadata sync (INDEX, manifest) is a separate step per spec.
Do not overwrite SKILL.md with this skill; curate-skills updates agent.yaml and README per skill. SKILL.md remains the canonical definition per spec.
INDEX.md is the canonical capability list (registry, tags, version, purpose); do not overwrite it. ASQM_AUDIT.md is the repo-level curation artifact: quality, lifecycle, overlaps, ecosystem, findings, and final recommendations (actionable next steps or “no changes”); write or update it on full curation runs and commit it.
Respect existing tags from skills/INDEX.md when normalizing; add or suggest tags only when clearly aligned with the tagging system.
Strict scoring: Apply ASQM dimensions strictly; do not inflate scores. agent_native = 5 only when the skill has an explicit output contract (Appendix or equivalent) in SKILL.md.

不得通过本Skill修改spec/skill.md或manifest.json；元数据同步（INDEX、manifest）是独立的标准化步骤。
不得通过本Skill覆盖SKILL.md；Skill治理仅更新每个Skill的agent.yaml与README。SKILL.md仍是标准化的权威定义文件。
INDEX.md是权威的能力列表（注册表、标签、版本、目标）；不得覆盖。ASQM_AUDIT.md是仓库级治理产物：包含质量、生命周期、重叠情况、生态系统、审计发现、最终建议（可执行步骤或“无需修改”）；在全量治理时生成或更新并提交。
标准化时需遵循skills/INDEX.md中已有的标签；仅当与标签体系高度匹配时，才可添加或建议新标签。
严格打分：严格应用ASQM维度规则；不得虚高分数。仅当SKILL.md中存在明确的输出契约（附录或等效内容）时，agent_native才可打5分。

Self-Check

自我检查

Examples

示例

Example 1: Full curation run

示例1：全量治理执行

Input:
```
skills_directory: skills/
```
; user said “curate all skills in this repo.”
Expected: Scan all subdirs of
```
skills/
```
; read each skill’s agent.yaml or README/SKILL.md; score; assign overlaps_with (owner/repo:skill-name) and market_position per skill; write back agent.yaml and normalized README; report overlaps and market_position; write ASQM_AUDIT.md or print structured summary. Confirm once before writing if policy applies.

输入：
```
skills_directory: skills/
```
；用户指令“治理本仓库中所有Skill”。
预期结果：扫描
```
skills/
```
下的所有子目录；读取每个Skill的agent.yaml或文档；打分；为每个Skill分配overlaps_with（owner/repo:skill-name）与market_position；回写agent.yaml与标准化README；报告重叠与定位情况；生成ASQM_AUDIT.md或打印结构化摘要。若交互规则要求，需先确认再执行。

Example 2: Single-skill re-score

示例2：单个Skill重新打分

Input: User says “re-score and update only refine-skill-design.”
Expected: Read that skill’s agent.yaml or docs; compute scores, status, overlaps_with, and market_position; update only that skill’s agent.yaml and README; do not write ASQM_AUDIT.md unless requested.

输入：用户指令“仅重新打分并更新refine-skill-design”。
预期结果：读取该Skill的agent.yaml或文档；计算分数、状态、overlaps_with与market_position；仅更新该Skill的agent.yaml与README；除非明确要求，否则不生成ASQM_AUDIT.md。

Edge case: New skill with no agent.yaml

边缘场景：无agent.yaml的新Skill

Input: A new skill directory has only SKILL.md (no agent.yaml, no README).
Expected: Read SKILL.md; derive scores, overlaps_with, and market_position; create agent.yaml with scores, status, overlaps_with, and market_position; generate a minimal normalized README from SKILL.md. Report in summary that the skill was newly instrumented.

输入：新Skill目录仅包含SKILL.md（无agent.yaml、无README）。
预期结果：读取SKILL.md；推导分数、overlaps_with与market_position；创建包含分数、状态、overlaps_with与market_position的agent.yaml；从SKILL.md生成最简标准化README。在摘要中说明该Skill为新增并已完成配置。

Appendix: Output contract (agent.yaml per skill)

附录：输出契约（单Skill的agent.yaml）

When this skill writes or updates a skill’s

agent.yaml

, it uses this structure so Agents can consume it without reading README:

yaml

name: [kebab-case skill name]
status: validated | experimental | archive_candidate

primary_use: [one-line purpose]

inputs:
  - [list of input names]

outputs:
  - [list of output artifacts]

scores:
  agent_native: [0-5]
  cognitive: [0-5]
  composability: [0-5]
  stance: [0-5]

asqm_quality: [0-20, linear: agent_native + cognitive + composability + stance]

overlaps_with:   # Git-repo form: owner/repo:skill-name (this repo and others alike)
  - [owner/repo:skill-name]
  - [owner/repo:other-skill]

market_position: differentiated | commodity | experimental

Scores (ASQM) measure intrinsic quality. asqm_quality = agent_native + cognitive + composability + stance (0–20). Lifecycle: validated ↔ Quality ≥ 17 AND agent_native ≥ 4 AND stance ≥ 3; experimental ↔ Quality ≥ 10; archive_candidate ↔ otherwise.
overlaps_with lists overlapping skills in Git-repo form
```
owner/repo:skill-name
```
(same format for this repo and other repos); empty
```
[]
```
when none.

当本Skill编写或更新Skill的

agent.yaml

时，采用以下结构，确保Agent无需读取README即可解析：

yaml

name: [kebab格式的Skill名称]
status: validated | experimental | archive_candidate

primary_use: [单行目标描述]

inputs:
  - [输入名称列表]

outputs:
  - [输出产物列表]

scores:
  agent_native: [0-5]
  cognitive: [0-5]
  composability: [0-5]
  stance: [0-5]

asqm_quality: [0-20，线性求和：agent_native + cognitive + composability + stance]

overlaps_with:   # Git仓库格式：owner/repo:skill-name（本仓库与外部仓库格式一致）
  - [owner/repo:skill-name]
  - [owner/repo:other-skill]

market_position: differentiated | commodity | experimental

ASQM评分：衡量内在质量。asqm_quality = agent_native + cognitive + composability + stance（0-20分）。生命周期：已验证 ↔ 质量分≥17 且 agent_native≥4 且 stance≥3；实验性 ↔ 质量分≥10；待归档候选 ↔ 其余情况。
overlaps_with：重叠Skill列表，采用Git仓库格式
```
owner/repo:skill-name
```
（本仓库与外部仓库格式统一）；无重叠则为
```
[]
```
。