AgentSkillsCN

verification-before-completion

强制代理在声称完成任务之前,先执行验证命令并出示证据。杜绝“应该能行”却缺乏实际证明的情况。

SKILL.md
--- frontmatter
name: verification-before-completion
description: Forces agents to run verification commands and show evidence before claiming completion. Prevents "it should work" without proof.

Verification Before Completion

Load this skill when you need to ensure work is actually verified, not just claimed complete.

Core Principle

NOTHING is "done" without PROOF it works.

Claims without evidence are worthless. Every completion claim must be backed by:

  • Command output showing success
  • Test results showing pass
  • Build output showing no errors
  • Observable behavior matching requirements

Verification Checklist

Before marking ANY task complete, verify:

1. Type Safety

bash
# Run LSP diagnostics on changed files
lsp_diagnostics(filePath="path/to/changed/file.ts")

Required Evidence: Clean output (no errors)

2. Build Success

bash
# Run project build command
bun run build  # or npm run build, etc.

Required Evidence: Exit code 0, no error messages

3. Test Passage

bash
# Run relevant tests
bun test  # or npm test, pytest, etc.

Required Evidence: All tests pass (or document pre-existing failures)

4. Manual Verification

  • Read the changed files
  • Verify changes match requirements
  • Check edge cases

Required Evidence: Describe what you observed


Anti-Patterns (BLOCKING)

ViolationWhy It Fails
"It should work now"No evidence. Run it.
"I added the tests"Did they pass? Show output.
"Fixed the bug"How do you know? What did you test?
"Implementation complete"Did you verify against success criteria?
Skipping test executionTests exist to be RUN
"The code looks correct"Looking isn't testing. Execute it.

Evidence Format

When reporting completion, include:

code
## Verification Evidence

### Build
✅ `bun run build` - Exit code 0
Output: [paste relevant output]

### Tests
✅ `bun test` - 42 tests passed
Output: [paste test summary]

### Type Check
✅ `lsp_diagnostics` - No errors in changed files

### Manual Check
✅ Verified [specific behavior] works as expected

When to Load This Skill

  • Before completing any implementation task
  • When delegating to coder/frontend agents
  • Before reporting task completion to user
  • When reviewing subagent work

Quick Reference

PhaseCommandEvidence Needed
Buildbun run buildExit code 0
Testbun testAll pass
Typeslsp_diagnosticsNo errors
Lintbun run lintNo warnings

CLAIM NOTHING WITHOUT PROOF. EXECUTE. VERIFY. SHOW EVIDENCE.