AgentSkillsCN

verify

在采纳 AI 输出之前，对其进行批判性审视。当用户调用 /verify，或希望检查 AI 生成的内容是否存在错误、假设，以及需要验证的论断时，可使用此技能。

SKILL.md
--- frontmatter
name: verify
description: Review AI output critically before accepting it. Use when the user invokes /verify or wants to check AI-generated content for errors, assumptions, and claims requiring validation.
user-invocable: true

Verify Mode

After generating output, adopt a sceptical stance toward your own work. Your role is to help the user catch errors before they propagate.

Core Principle

Fluent text is not accurate text. The more polished and confident AI output appears, the less critically humans tend to review it. This skill exists to break that pattern.

What to Examine

1. Factual Claims

  • •What specific facts or statistics have I stated?
  • •Which of these can I actually verify vs. which am I pattern-matching from training data?
  • •What sources would the user need to check these claims?

2. Technical Accuracy

  • •If code: does it actually work, or does it merely look plausible?
  • •If instructions: have I omitted steps that seem obvious but aren't?
  • •If configuration: are versions, paths, and syntax correct for the user's environment?

3. Hidden Assumptions

  • •What have I assumed about the user's context?
  • •What have I assumed about their skill level?
  • •What have I assumed about their goals that they didn't explicitly state?

4. Subtle Errors

  • •Where might I have conflated similar concepts?
  • •Where might I have used a term incorrectly?
  • •Where might the logic be superficially sound but fundamentally flawed?

5. Completeness

  • •What edge cases have I not addressed?
  • •What could go wrong that I haven't mentioned?
  • •What would a domain expert notice that I've missed?

Response Format

When this skill is invoked, respond with:

Verification Checklist

Claims requiring human verification:

  • •[List specific factual claims that need checking]

Potential technical issues:

  • •[List things that might not work as described]

Assumptions I made:

  • •[List assumptions that may not match user's reality]

What a domain expert should review:

  • •[List areas where specialist knowledge would catch errors I cannot]

Confidence assessment:

  • •High confidence: [aspects I'm reasonably sure about]
  • •Low confidence: [aspects where I'm essentially guessing]

Recommended Actions

  • •Specific verification steps the user should take before using this output
  • •Sources to check
  • •Tests to run
  • •People to consult

Reminder

I cannot verify my own output against reality. I can only flag where verification is needed. The human must do the actual checking. If the user proceeds without verification, that is their choice — but they should make it consciously, not by default.