## Success Criteria

### Context Gathering (20 points)
- Agent extracts all relevant context from screen description
- Visible elements are listed with their states
- System information is captured (application, version, mode)
- User instruction is preserved in its original form

### Instruction Parsing (20 points)
- Action verb is identified correctly
- Target is specified with enough detail to locate
- Modifier adds specificity (how, where, when)
- Scope is defined (one instance vs all instances)

### Command Precision (25 points)
- Command has a clear action verb (create, modify, fix, add, remove)
- Target is specific (file path, function name, line number)
- At least 2 precision clauses are included (location, format, constraint, quality)
- Constraints prevent unintended side effects
- Completion criteria are concrete and testable

### Ambiguity Handling (15 points)
- Vague instructions are detected
- Clarifying questions are asked when needed
- Proposed interpretation is offered for confirmation
- Agent does not guess when critical information is missing

### Command Format (10 points)
- Command follows a structured format
- Task summary is one line
- Details are listed as bullet points
- Constraints are separate from details
- Completion criteria are clearly stated

### Anti-Pattern Avoidance (10 points)
- No vague verbs without specifics
- No missing targets
- No assumed context
- No kitchen-sink commands (one command trying to do too many things)

## Evaluation Scenarios

### Scenario A: Vague Input, No Context
- User says "fix it" with no screen context
- Agent asks for clarification with specific questions
- **Result:** PASS — correctly identifies ambiguity

### Scenario B: Vague Input, Rich Context
- User says "add error handling" with screen showing specific code
- Agent produces precise command targeting the visible code
- **Result:** PASS — uses context to resolve ambiguity

### Scenario C: Specific Input, No Context
- User says "add phone_number column to users table"
- Agent produces command but asks about nullable/required
- **Result:** PASS — identifies missing details

### Scenario D: Agent Guesses Without Asking
- User says "fix the button"
- Agent assumes it is a CSS issue and modifies styles
- Button was actually broken due to JavaScript error
- **Result:** FAIL — guessed instead of clarifying

### Scenario E: Kitchen Sink Command
- User says "clean up the file"
- Agent produces a single command with 10 different changes
- **Result:** FAIL — too many tasks in one command

## Red Flags (Automatic Deductions)
- Command issued without a clear target: -10 points
- Vague verb used without specificity: -5 points per instance
- No precision clauses included: -10 points
- Agent guesses when instruction is ambiguous: -15 points
- Completion criteria are vague ("make it better"): -10 points
- Command tries to do more than 3 unrelated things: -10 points