Verification-First Development

The Problem with AI Coding Assistants

Most AI coding assistants:

• Claim fixes without verification
• Move on before testing
• Celebrate partial progress as complete
• Avoid acknowledging broken code

This System's Approach

Rule 1: Prove It Works

Every code change MUST be verified before moving on:

• Run the build
• Run the tests
• Check the actual output
• Confirm expected behavior

Rule 2: Honest Failure Reporting

When something fails:

• State clearly: "This failed"
• Include the actual error message
• Don't soften or interpret the failure
• Don't say "almost working" - it either works or it doesn't

Rule 3: No Assumptions

• Don't assume a fix worked
• Don't assume tests pass
• Don't assume code compiles
• VERIFY EVERYTHING

Rule 4: Root Cause First

Before attempting any fix:

1. Read the actual error message
2. Understand what the error means
3. Identify the root cause
4. THEN propose a fix

Rule 5: One Fix at a Time

• Make one change
• Verify it
• Then move to the next
• Don't stack multiple untested changes

Verification Commands

Frontend (Next.js)

# Type check
pnpm type-check

# Lint
pnpm lint

# Build
pnpm build

# Test
pnpm test

Backend (Python)

# Type check
uv run mypy src/

# Lint
uv run ruff check src/

# Test
uv run pytest

# Run server
uv run uvicorn src.api.main:app --reload

Full Stack

# Everything
pnpm turbo run build lint type-check test

Australian Context Verification

Additional verification for Australian market:

• Spelling: Verify en-AU spelling (colour, organisation, licence)
• Dates: Verify DD/MM/YYYY format in UI
• Currency: Verify AUD currency formatting ($1,234.56)
• Regulations: Verify compliance with Privacy Act 1988, WCAG 2.1 AA
• Design: Verify NO Lucide icons used, 2025-2026 aesthetic enforced

Verification Tiers

Tier A: Quick Verification (30 seconds)

• Type check passes
• Lint passes
• No console errors

Tier B: Standard Verification (2-3 minutes)

• All Tier A checks
• Unit tests pass
• Build succeeds
• Manual smoke test of changed functionality

Tier C: Full Verification (5-10 minutes)

• All Tier B checks
• Integration tests pass
• E2E tests for affected flows
• Visual regression check
• Performance benchmark

Tier D: Production Verification (15-20 minutes)

• All Tier C checks
• Full test suite (including slow tests)
• Lighthouse audit >90 all scores
• Security scan
• Accessibility audit (WCAG 2.1 AA)
• Australian context verification

Output Format

When reporting status, use this format:

## Task: [Description]

### Status: [PASS | FAIL | BLOCKED]

### Verification Tier: [A | B | C | D]

### Verification:
- Build: [PASS/FAIL] - [output if failed]
- Tests: [PASS/FAIL] - [X/Y passed, failures listed]
- Manual check: [PASS/FAIL] - [what was checked]
- Australian context: [PASS/FAIL/N/A] - [en-AU compliance]

### Evidence:
[Screenshots, test output, logs]

### Next Steps:
- [If PASS: what's next]
- [If FAIL: what needs to be fixed and why]

Integration with Hooks

This skill is automatically enforced by:

• post-code.hook.md - Runs Tier A verification after code generation
• pre-commit.hook.md - Runs Tier B verification before git commit
• pre-deploy.hook.md - Runs Tier D verification before deployment

Independent Verification

Critical: Verification MUST be independent. The verification agent (.claude/agents/verification/) performs verification, NOT the agent that wrote the code.

NO SELF-ATTESTATION. Evidence required.