Review Gates

source-to-skill should not only generate skill artifacts. It should protect the skill system from attractive but weak updates.

The long-term rule is:

source -> delta -> review gates -> user decision -> skill evolution

A source should become a new skill only when it survives the same pressure that would be applied to an update of an existing skill.

When To Run The Gates

Run these gates when a source might:

• create a Mini Skill or Full Skill
• change ## Core Guidance in an existing skill
• add a new decision rule, workflow, checklist, or framework
• resolve or introduce a contradiction
• change when a skill should or should not be used

The gates are lighter for notes, seeds, and evidence-only updates. They are strict for anything that changes agent behavior.

Gate 1: First-Principles Review

The first-principles review asks whether the proposed guidance still serves the core job of the skill.

Review questions:

Question	Pass signal	Fail signal
What job should this skill help an agent do?	The job is concrete and task-shaped	The job is a vague topic or content category
What is the core method?	The method has steps, checks, or decisions	The method is only a theme or opinion
Does the new source strengthen that method?	It clarifies, narrows, or supports the method	It adds attractive but unrelated content
Is the method transferable?	It can work across sources, authors, or cases	It only describes one local anecdote
What is the boundary?	The skill says when not to use the method	The skill implies universal usefulness

Chinese shorthand:

第一性原理测试：
这个 skill 到底帮 agent 做什么？
这条规则是否真的服务这个核心任务？
它是方法的一部分，还是只是一个好看的观点？
换一个 source / 作者 / 场景后，它还能成立吗？

Gate 2: Adversarial Review

The adversarial review tries to break the proposed guidance before it becomes persistent behavior.

Review questions:

Attack	Question	Required response
Counterexample	What example would make this rule fail?	Add a boundary, exception, or reject the rule
Contradiction	Does this conflict with existing guidance?	Ask the user or record a contradiction
Overgeneralization	Is a local case being turned into a universal rule?	Narrow the claim
Evidence gap	Did the source actually support this claim?	Add evidence or delete the claim
Use-case drift	Does this make the skill solve a different problem?	Split, reject, or create a separate seed
Unsafe merge	Could this update make the agent worse in common tasks?	Keep as pending until reviewed

Chinese shorthand:

对抗式审查：
有什么反例会击穿这条规则？
它和已有 skill 有没有冲突？
它是不是把局部案例包装成通用方法？
证据是否足够，还是 AI 自己补全了作者没说的话？
适用边界、失败条件、不要使用的场景是否清楚？

Relationship Decisions

After both gates, classify the source relationship:

Decision	Use when	Action
duplicate	The source repeats known guidance	Do not update core guidance
evidence	The source supports an existing rule	Add to evidence only
refinement	The source improves precision	Propose a targeted edit
contradiction	The source conflicts with a rule	Ask the user before merging
new_skill	The source has a stable method outside existing skills	Create a new skill
reject	The source is noisy, unsupported, or misleading	Record or discard

User Review Gate

Do not ask the user to review everything. Ask only at decision points:

• accepting a contradiction
• rewriting core guidance
• creating a new skill instead of updating an old one
• promoting a seed into a Mini Skill or Full Skill
• changing use_when / do_not_use_when

Useful prompt shape:

This source overlaps with `product-positioning-skill`.

Recommended action: refinement

Conflict:
- Existing guidance: early products should start with narrow positioning.
- New source: early products may need broad public attention before narrowing.

Suggested merge:
- Treat early positioning as a temporary expression layer, not a fixed strategy.

Choose:
1. merge as refinement
2. keep as contradiction
3. add as evidence only
4. reject
5. create a separate skill seed

Review Report Template

Use this shape for future evolution/pending-updates/ artifacts:

# Skill Update Review

Source:
Target skill:
Recommended relationship:

## Extracted Method Candidates

## First-Principles Review

Core job:
Core method:
Transferability:
Boundary:
Pass / fail:

## Adversarial Review

Counterexamples:
Contradictions:
Overgeneralization risk:
Evidence gaps:
Use-case drift:
Unsafe merge risk:
Pass / fail:

## Proposed Change

## Evidence To Add

## User Decision Needed

## Final Decision

Completion Rule

A proposed update is not ready to merge until:

• each new guidance claim has evidence
• boundaries and failure cases are stated
• contradictions are reviewed by a human
• the target skill still has one clear job
• evaluation or regression questions have been updated when behavior changes

If any of these are missing, keep the update pending.