AI 工作资产化诊断器

中文速查

• 中文名：AI 工作资产化诊断器 / 资产化路由器
• 英文稳定名：ai-work-assetization-diagnoser
• 分类：Skill/Agent 治理
• 你可以这样叫我：这段工作是不是值得做成 Skill、这个 prompt 应该沉淀成 workflow 还是 Skill、帮我判断该资产化到哪层、这个 AI 工作流要不要做成 Loop
• 适合：判断一段可重复 AI 工作应该沉淀到哪个资产层，给出最小下一步 artifact 和验证信号。
• 不适合：直接创建 Skill、直接实现自动化系统、普通日志根因定位、一次性事实查询、没有复用价值的闲聊。

Overview

这个 Skill 是 Router / Gate，不是执行器。它回答三个问题：

1. 这段 AI 工作是否值得沉淀。
2. 如果值得，最小有用资产层是什么。
3. 为什么不是相邻层级。

默认输出应短、可执行、有证据。不要把所有重复任务都升级成 Skill，也不要把所有自动化想法都升级成 Loop。

Asset Layers

Layer	Use When	Example Artifact
Do Not Assetize	低频、一次性、强主观、输入不可稳定复用、风险高或验收口径不存在。	保留聊天记录或一次性笔记。
Prompt	步骤简单，主要复用表达方式。	Prompt template, checklist prompt.
Context Pack	关键难点是资料、约束、样例和反例组织。	Context folder, source bundle, glossary.
Workflow	有稳定步骤、角色、输入输出和人工推进点。	SOP, runbook, workflow doc.
Skill	高频可复用，有明确触发语、输入、输出、边界和验收方式。	SKILL.md + references/scripts.
Loop	需要多轮状态、恢复、触发器、检查点、重试或人工接管。	Loop contract, state files, update log.
System	多个 Skill/Loop/Agent 组合，涉及权限、成本、审计、评估或团队级运行。	Product/system PRD, architecture plan.

Workflow

1. Identify input shape

   • AI conversation
   • repeated manual task
   • prompt or prompt pack
   • team workflow
   • successful delivery trace
   • failed or over-engineered asset proposal

2. Extract evidence

   • user goal and business context
   • input materials and constraints
   • repeated steps
   • output artifact
   • human decision points
   • validation or acceptance criteria
   • frequency and reuse audience
   • state, retry, handoff, or automation needs

3. Score only what matters

   • repeatability
   • input stability
   • output stability
   • validation clarity
   • context dependency
   • human judgment dependency
   • failure cost
   • reuse audience

4. Recommend one primary asset layer

   • Give the smallest layer that would create real reuse.
   • Explain why the lower layer is insufficient.
   • Explain why the higher layer is overkill.
   • Include do-not-assetize if that is the best answer.

5. Name the smallest next artifact

   • Prompt: one reusable prompt with slots.
   • Context Pack: file list and required metadata.
   • Workflow: step list and handoff format.
   • Skill: trigger description, non-triggers, output contract, references.
   • Loop: state files, stop conditions, Human Gate.
   • System: PRD / architecture / eval plan before implementation.

6. Define reuse signal

   • What would prove this asset is worth deepening?
   • What failure signal should stop further investment?

Output Contract

Use this structure:

## Assetization Diagnosis

- Current shape: <conversation / prompt / context / workflow / skill candidate / loop candidate / system candidate>
- Recommended layer: <Do Not Assetize / Prompt / Context Pack / Workflow / Skill / Loop / System>
- Confidence: <High / Medium / Low>

## Evidence
- ...

## Why This Layer
- ...

## Why Not Adjacent Layers
- Lower layer is insufficient because ...
- Higher layer is overkill because ...

## Smallest Next Artifact
- ...

## Reuse Signal To Watch
- ...

## Do Not Do
- ...

For batch scenarios, use a compact table with columns:

• Scenario
• Current shape
• Recommended layer
• Evidence
• Smallest next artifact
• Reuse signal
• Priority

Boundary Rules

• If the user asks to create the actual Skill after diagnosis, hand off to skill-creator.
• If the user asks whether an idea is a good product, use ai-collaboration-calibration or decision-research.
• If the user asks to turn a product idea into a PRD, use prd-architect.
• If the user asks to review an existing Skill, use skill-reviewer when available.
• If the user provides trace/logs and wants root cause, use a trace/debug skill, not this one.
• If the work is high responsibility, such as medical, legal, financial, safety, hiring, or compliance decisions, recommend human review and avoid closed-loop automation.

Non-Assetization Cases

Recommend not assetizing, or only a lightweight note, when:

• The task happened once and is unlikely to recur.
• The input varies so much that templates would mislead.
• The output depends on taste or senior judgment with no stable rubric.
• The cost of a wrong answer is high and no verification exists.
• The user cannot name a reuse audience.
• A simpler checklist would solve the problem.

Definition Of Done

• The recommendation is one primary asset layer, not a vague list.
• Adjacent layers are explicitly accepted or rejected.
• The next artifact is small enough to create in one focused pass.
• The answer includes a reuse or failure signal.
• The diagnosis does not expose private context unless the user supplied and authorized it in this run.

Resources

• references/asset-layer-rubric.md：当 Prompt / Context Pack / Workflow / Skill / Loop / System 边界不清时读取。
• scripts/check_assetization_report.py：检查资产化诊断报告是否包含推荐层级、证据、最小下一步和复用信号。

Evaluation

Smoke prompts:

• 这段工作是不是值得做成 Skill？
• 这个 prompt 应该沉淀成 workflow 还是 Skill？
• 我们连续三次做了类似 AI 协作，帮我判断该资产化到哪层。
• 这只是一次性任务，不要过度沉淀。

Non-trigger prompts:

• 直接帮我创建一个 Skill。
• 看下这个 trace，定位根因。
• 帮我写一个 PRD。
• 查一下今天的新闻。