Evidence-Based Verification Skill

Version: 1.0.0 Type: Quality Assurance Auto-activate: Code review, task completion, production deployment

Overview

This skill teaches agents how to collect and verify evidence before marking tasks complete. Inspired by production-grade development practices, it ensures all claims are backed by executable proof: test results, coverage metrics, build success, and deployment verification.

Key Principle: Show, don't tell. No task is complete without verifiable evidence.

When to Use This Skill

Auto-Activate Triggers

• Completing code implementation
• Finishing code review
• Marking tasks complete in Squad mode
• Before agent handoff
• Production deployment verification

Manual Activation

• When user requests "verify this works"
• Before creating pull requests
• During quality assurance reviews
• When troubleshooting failures

Core Concepts

1. Evidence Types

Test Evidence

• Exit code (must be 0 for success)
• Test suite results (passed/failed/skipped)
• Coverage percentage (if available)
• Test duration

Build Evidence

• Build exit code (0 = success)
• Compilation errors/warnings
• Build artifacts created
• Build duration

Deployment Evidence

• Deployment status (success/failed)
• Environment deployed to
• Health check results
• Rollback capability verified

Code Quality Evidence

• Linter results (errors/warnings)
• Type checker results
• Security scan results
• Accessibility audit results

2. Evidence Collection Protocol

## Evidence Collection Steps

1. **Identify Verification Points**
   - What needs to be proven?
   - What could go wrong?
   - What does "complete" mean?

2. **Execute Verification**
   - Run tests
   - Run build
   - Run linters
   - Check deployments

3. **Capture Results**
   - Record exit codes
   - Save output snippets
   - Note timestamps
   - Document environment

4. **Store Evidence**
   - Add to shared context
   - Reference in task completion
   - Link to artifacts

3. Verification Standards

Minimum Evidence Requirements:

• ✅ At least ONE verification type executed
• ✅ Exit code captured (0 = pass, non-zero = fail)
• ✅ Timestamp recorded
• ✅ Evidence stored in context

Production-Grade Requirements:

• ✅ Tests run with exit code 0
• ✅ Coverage >70% (or project standard)
• ✅ Build succeeds with exit code 0
• ✅ No critical linter errors
• ✅ Security scan passes

Evidence Collection Templates

Template 1: Test Evidence

Use this template when running tests:

## Test Evidence

**Command:** `npm test` (or equivalent)
**Exit Code:** 0 ✅ / non-zero ❌
**Duration:** X seconds
**Results:**
- Tests passed: X
- Tests failed: X
- Tests skipped: X
- Coverage: X%

**Output Snippet:**

[First 10 lines of test output]

**Timestamp:** YYYY-MM-DD HH:MM:SS
**Environment:** Node vX.X.X, OS, etc.

Template 2: Build Evidence

Use this template when building:

## Build Evidence

**Command:** `npm run build` (or equivalent)
**Exit Code:** 0 ✅ / non-zero ❌
**Duration:** X seconds
**Artifacts Created:**
- dist/bundle.js (XXX KB)
- dist/styles.css (XXX KB)

**Errors:** X
**Warnings:** X

**Output Snippet:**

[First 10 lines of build output]

**Timestamp:** YYYY-MM-DD HH:MM:SS

Template 3: Code Quality Evidence

Use this template for linting and type checking:

## Code Quality Evidence

**Linter:** ESLint / Ruff / etc.
**Command:** `npm run lint`
**Exit Code:** 0 ✅ / non-zero ❌
**Errors:** X
**Warnings:** X

**Type Checker:** TypeScript / mypy / etc.
**Command:** `npm run typecheck`
**Exit Code:** 0 ✅ / non-zero ❌
**Type Errors:** X

**Timestamp:** YYYY-MM-DD HH:MM:SS

Template 4: Combined Evidence Report

Use this comprehensive template for task completion:

## Task Completion Evidence

### Task: [Task description]
### Agent: [Agent name]
### Completed: YYYY-MM-DD HH:MM:SS

### Verification Results

| Check | Command | Exit Code | Result |
|-------|---------|-----------|--------|
| Tests | `npm test` | 0 | ✅ 45 passed, 0 failed |
| Build | `npm run build` | 0 | ✅ Bundle created (234 KB) |
| Linter | `npm run lint` | 0 | ✅ No errors, 2 warnings |
| Types | `npm run typecheck` | 0 | ✅ No type errors |

### Coverage
- Statements: 87%
- Branches: 82%
- Functions: 90%
- Lines: 86%

### Evidence Files
- Test output: `.claude/quality-gates/evidence/tests-2025-XX-XX.log`
- Build output: `.claude/quality-gates/evidence/build-2025-XX-XX.log`

### Conclusion
All verification checks passed. Task ready for review.

Step-by-Step Workflows

Workflow 1: Code Implementation Verification

When: After writing code for a feature or bug fix

Steps:

1. Save all files - Ensure changes are written

2. Run tests

   bashCopy

   npm test
   # or: pytest, cargo test, go test, etc.
   

   • Capture exit code
   • Note passed/failed counts
   • Record coverage if available

3. Run build (if applicable)

   bashCopy

   npm run build
   # or: cargo build, go build, etc.
   

   • Capture exit code
   • Note any errors/warnings
   • Verify artifacts created

4. Run linter

   bashCopy

   npm run lint
   # or: ruff check, cargo clippy, golangci-lint run
   

   • Capture exit code
   • Note errors/warnings

5. Run type checker (if applicable)

   bashCopy

   npm run typecheck
   # or: mypy, tsc --noEmit
   

   • Capture exit code
   • Note type errors

6. Document evidence

   • Use Template 4 (Combined Evidence Report)
   • Add to shared context under quality_evidence
   • Reference in task completion message

7. Mark task complete (only if all evidence passes)

Workflow 2: Code Review Verification

When: Reviewing another agent's code or user's PR

Steps:

1. Read the code changes

2. Verify tests exist

   • Are there tests for new functionality?
   • Do tests cover edge cases?
   • Are existing tests updated?

3. Run tests

   • Execute test suite
   • Verify exit code 0
   • Check coverage didn't decrease

4. Check build

   • Ensure project still builds
   • No new build errors

5. Verify code quality

   • Run linter
   • Run type checker
   • Check for security issues

6. Document review evidence

   • Use Template 3 (Code Quality Evidence)
   • Note any issues found
   • Add to context

7. Approve or request changes

   • Approve only if all evidence passes
   • If issues found, document them with evidence

Workflow 3: Production Deployment Verification

When: Deploying to production or staging

Steps:

1. Pre-deployment checks

   • All tests pass (exit code 0)
   • Build succeeds
   • No critical linter errors
   • Security scan passes

2. Execute deployment

   • Run deployment command
   • Capture output

3. Post-deployment checks

   • Health check endpoint responds
   • Application starts successfully
   • No immediate errors in logs
   • Smoke tests pass

4. Document deployment evidence

   markdownCopy

   ## Deployment Evidence
   
   **Environment:** production
   **Timestamp:** YYYY-MM-DD HH:MM:SS
   **Version:** vX.X.X
   
   **Pre-Deployment:**
   - Tests: ✅ Exit 0
   - Build: ✅ Exit 0
   - Security: ✅ No critical issues
   
   **Deployment:**
   - Command: `kubectl apply -f deployment.yaml`
   - Exit Code: 0 ✅
   
   **Post-Deployment:**
   - Health Check: ✅ 200 OK
   - Smoke Tests: ✅ All passed
   - Error Rate: <0.1%
   

5. Verify rollback capability

   • Ensure previous version can be restored
   • Document rollback procedure

Evidence Storage

Where to Store Evidence

Shared Context (Primary)

{
  "quality_evidence": {
    "tests_run": true,
    "test_exit_code": 0,
    "coverage_percent": 87,
    "build_success": true,
    "build_exit_code": 0,
    "linter_errors": 0,
    "linter_warnings": 2,
    "timestamp": "2025-11-02T10:30:00Z"
  }
}

Evidence Files (Secondary)

• .claude/quality-gates/evidence/ directory
• One file per verification run
• Format: {type}-{timestamp}.log
• Example: tests-2025-11-02-103000.log

Task Completion Messages

• Include evidence summary
• Link to detailed evidence files
• Example: "Task complete. Tests passed (exit 0, 87% coverage), build succeeded."

Quality Standards

Minimum Acceptable

✅ Tests executed with captured exit code ✅ Timestamp recorded ✅ Evidence stored in context

Production-Grade

✅ Tests pass (exit code 0) ✅ Coverage ≥70% (or project standard) ✅ Build succeeds (exit code 0) ✅ No critical linter errors ✅ Type checker passes ✅ Security scan shows no critical issues

Gold Standard

✅ All production-grade requirements ✅ Coverage ≥80% ✅ No linter warnings ✅ Performance benchmarks within thresholds ✅ Accessibility audit passes (WCAG 2.1 AA) ✅ Integration tests pass ✅ Deployment verification complete

Common Pitfalls

❌ Don't Skip Evidence Collection

Bad:

"I've implemented the login feature. It should work correctly."

Good:

"I've implemented the login feature. Evidence:
- Tests: Exit code 0, 12 tests passed, 0 failed
- Build: Exit code 0, no errors
- Coverage: 89%
Task complete with verification."

❌ Don't Fake Evidence

Bad:

"Tests passed" (without actually running them)

Good:

"Tests passed. Exit code: 0
Command: npm test
Output: Test Suites: 3 passed, 3 total
Timestamp: 2025-11-02 10:30:15"

❌ Don't Ignore Failed Evidence

Bad:

"Build failed with exit code 1, but the code looks correct so marking complete."

Good:

"Build failed with exit code 1. Errors:
- TypeError: Cannot read property 'id' of undefined (line 42)
Fixing the error now before marking complete."

❌ Don't Collect Evidence Only Once

Bad:

"Tests passed yesterday, so the code is still good."

Good:

"Re-running tests after today's changes.
New evidence: Exit code 0, 45 tests passed, coverage 87%"

Integration with Other Systems

Context System Integration

Evidence is automatically tracked in shared context:

// Context structure includes:
{
  quality_evidence?: {
    tests_run: boolean;
    test_exit_code?: number;
    coverage_percent?: number;
    build_success?: boolean;
    linter_errors?: number;
    timestamp: string;
  }
}

Quality Gates Integration

Evidence collection feeds into quality gates:

• Quality gates check if evidence exists
• Block task completion if evidence missing
• Escalate if evidence shows failures

Squad Mode Integration

In parallel execution:

• Each agent collects evidence independently
• Studio Coach validates evidence before sync
• Blocked tasks don't waste parallel cycles

Quick Reference

Evidence Collection Checklist

Before marking task complete:

- [ ] Tests executed
- [ ] Test exit code captured (0 = pass)
- [ ] Build executed (if applicable)
- [ ] Build exit code captured (0 = pass)
- [ ] Code quality checks run (linter, types)
- [ ] Evidence documented with timestamp
- [ ] Evidence added to shared context
- [ ] Evidence summary in completion message

Common Commands by Language/Framework

JavaScript/TypeScript:

npm test                 # Run tests
npm run build           # Build project
npm run lint            # Run ESLint
npm run typecheck       # Run TypeScript compiler

Python:

pytest                  # Run tests
pytest --cov           # Run tests with coverage
ruff check .           # Run linter
mypy .                 # Run type checker

Rust:

cargo test             # Run tests
cargo build            # Build project
cargo clippy           # Run linter

Go:

go test ./...          # Run tests
go build               # Build project
golangci-lint run      # Run linter

Examples

See /skills/evidence-verification/examples/ for:

• Sample evidence reports
• Real-world verification scenarios
• Integration examples

Version History

v1.0.0 - Initial release

• Core evidence collection templates
• Verification workflows
• Quality standards
• Integration with context system

Remember: Evidence-first development prevents hallucinations, ensures production quality, and builds confidence. When in doubt, collect more evidence, not less.