ai-dlc-completion-criteria

# AI-DLC Completion Criteria

Completion criteria are the foundation of effective AI-DLC workflows. Clear criteria enable autonomy, provide exit conditions, and ensure alignment between human intent and AI execution.

## Why Criteria Matter

### The Autonomy Equation

```
Autonomy = f(Criteria Clarity)
```

- **Vague criteria** → Constant human oversight required
- **Clear criteria** → AI can self-verify and operate autonomously

### The Exit Condition Problem

Without clear criteria:
- When is the task "done"?
- How do we know if it's working?
- What constitutes success vs failure?

With clear criteria:
- Each criterion is a checkpoint
- All criteria satisfied = task complete
- Any criterion failed = more work needed

## Characteristics of Good Criteria

### 1. Specific

Bad: "Make the login work"
Good: "Users can log in with email and password"

Bad: "Improve performance"
Good: "API response time under 200ms for 95th percentile"

Bad: "Fix the bug"
Good: "Clicking 'Submit' no longer throws a TypeError"

### 2. Measurable

Can you write a test for it? If not, it's not measurable.

Bad: "The code should be clean"
Good: "No ESLint errors or warnings"

Bad: "The UI should be intuitive"
Good: "User can complete checkout in under 3 clicks"

Bad: "Should handle errors gracefully"
Good: "Invalid email shows 'Please enter a valid email' message"

### 3. Atomic

One criterion, one thing to check.

Bad: "Users can log in and their session persists"
Good:
- "Users can log in with valid credentials"
- "Session persists across page refreshes"

Bad: "API handles all edge cases"
Good:
- "Returns 400 for missing required fields"
- "Returns 404 for non-existent resources"
- "Returns 429 when rate limit exceeded"

### 4. Automated (Ideally)

Criteria that can be checked automatically are strongest.

```markdown
## Completion Criteria

- [ ] All unit tests pass (`bun test`)
- [ ] No TypeScript errors (`tsc --noEmit`)
- [ ] No lint warnings (`biome check`)
- [ ] API responds in < 200m