AgentSkillsCN

define-validation

此技能为正在进行中的工作定义AI验证标准。适用于在规格或计划中尚未明确验证标准，或在缺乏正式文档的情况下开始工作的场景。触发短语包括“AI将如何进行验证”、“定义验证标准”、“AI应该检查什么”、“针对这项工作进行验证”。

SKILL.md
--- frontmatter
name: define-validation
description: This skill defines AI validation criteria for work in progress. Use when validation wasn't defined in spec/plan, or when starting work without formal documentation. Triggers include "how will the AI validate", "define validation criteria", "what should the AI check", "validation for this work".

Define Validation

Define how the AI validates its work before declaring done.

When to Use

  • •Spec or plan exists but lacks AI Validation section
  • •Starting work without formal spec/plan
  • •Want to make validation criteria explicit for any chunk of work
  • •Reviewing existing criteria for completeness

Process

  1. •Identify the work: Read any existing spec, plan, or gather context from conversation
  2. •Start with defaults: Always include the standard validation checklist
  3. •Probe for custom needs: Ask "Does this feature need any specific verification beyond the defaults?"
  4. •Output the criteria: Present for user confirmation
  5. •Save or append: Either update existing spec/plan or save standalone

Output

If a spec or plan exists, offer to append the AI Validation section to it.

If no formal document exists, save to .lore/validation/[feature-or-work].md

Validation Criteria Structure

markdown
## AI Validation

**Defaults** (apply unless overridden):
- Unit tests with mocked time/network/filesystem/LLM calls (including Agent SDK `query()`)
- 90%+ coverage on new code
- Code review by fresh-context sub-agent

**Custom**:
- [Feature-specific validation steps]

Defaults Explained

These apply to virtually all work:

DefaultWhy
Mock timeTests shouldn't depend on when they run
Mock networkTests shouldn't fail due to connectivity
Mock filesystemTests should be isolated and reproducible
Mock LLM callsAgent SDK query() is an external API, costs money, can fail
90%+ coverageNew code should be exercised by tests
Code reviewFresh-context sub-agent catches what the implementer misses

Custom Validation Examples

When probing for custom needs, consider:

  • •CLI tools: "Output matches expected format in examples/"
  • •Parsers: "All test fixtures parse without errors"
  • •Generators: "Generated files are syntactically valid"
  • •Integrations: "Integration test passes against staging/mock API"
  • •UI components: "Renders without console errors in test harness"
  • •Data migrations: "Round-trip preserves data integrity"

Standalone Document Structure

When no spec/plan exists:

markdown
# Validation: [Work Description]

**For**: Brief description of what's being built

## AI Validation

**Defaults** (apply unless overridden):
- Unit tests with mocked time/network/filesystem/LLM calls (including Agent SDK `query()`)
- 90%+ coverage on new code
- Code review by fresh-context sub-agent

**Custom**:
- [Feature-specific items]

## Context
How this validation criteria was derived (conversation, informal description, etc.)

Keep It Actionable

Validation criteria must be things the AI can actually do:

  • •"Run the test suite" - actionable
  • •"Verify the user experience is good" - not actionable
  • •"Check output matches examples/expected.json" - actionable
  • •"Ensure performance is acceptable" - not actionable (unless threshold defined)