Skill("custom-plugin-prompt-engineering:prompt-evaluation")

parameters:
  evaluation_type:
    type: enum
    values: [accuracy, consistency, robustness, efficiency, ab_test]
    required: true

  test_cases:
    type: array
    items:
      input: string
      expected: string|object
      category: string
    min_items: 5

  metrics:
    type: array
    items: string
    default: [accuracy, consistency]

test_case:
  id: "TC001"
  category: "happy_path|edge_case|adversarial|regression"
  description: "What this test verifies"
  priority: "critical|high|medium|low"

  input:
    user_message: "Test input text"
    context: {}  # Optional additional context

  expected:
    contains: ["required", "phrases"]
    not_contains: ["forbidden", "content"]
    format: "json|text|markdown"
    schema: {}  # Optional JSON schema

  evaluation:
    metrics: ["accuracy", "format_compliance"]
    threshold: 0.9
    timeout: 30

categories:
  happy_path:
    description: "Standard expected inputs"
    coverage_target: 40%
    examples:
      - typical_user_query
      - common_use_case
      - expected_format

  edge_cases:
    description: "Boundary conditions"
    coverage_target: 25%
    examples:
      - empty_input
      - very_long_input
      - special_characters
      - unicode_text
      - minimal_input

  adversarial:
    description: "Attempts to break prompt"
    coverage_target: 20%
    examples:
      - injection_attempts
      - conflicting_instructions
      - ambiguous_requests
      - malformed_input

  regression:
    description: "Previously failed cases"
    coverage_target: 15%
    examples:
      - fixed_bugs
      - known_edge_cases

accuracy_rubric:
  1.0: "Exact match to expected output"
  0.8: "Semantically equivalent, minor differences"
  0.5: "Partially correct, major elements present"
  0.2: "Related but incorrect"
  0.0: "Completely wrong or off-topic"

consistency_rubric:
  1.0: "Identical across all runs"
  0.8: "Minor variations (punctuation, formatting)"
  0.5: "Significant variations, same meaning"
  0.2: "Different approaches, inconsistent"
  0.0: "Completely different each time"

robustness_rubric:
  1.0: "Handles all edge cases correctly"
  0.8: "Fails gracefully with clear error messages"
  0.5: "Some edge cases cause issues"
  0.2: "Many edge cases fail"
  0.0: "Breaks on most edge cases"

ab_test_config:
  name: "Prompt Variant Comparison"
  hypothesis: "New prompt improves accuracy by 10%"

  variants:
    control:
      prompt: "{original_prompt}"
      allocation: 50%
    treatment:
      prompt: "{new_prompt}"
      allocation: 50%

  metrics:
    primary:
      - name: accuracy
        min_improvement: 0.05
        significance: 0.05
    secondary:
      - token_count
      - response_time
      - user_satisfaction

  sample:
    minimum: 100
    power: 0.8

  stopping_rules:
    - condition: significant_regression
      action: stop_treatment
    - condition: clear_winner
      threshold: 0.99
      action: early_stop

report:
  metadata:
    prompt_name: "string"
    version: "string"
    date: "ISO8601"
    evaluator: "string"

  summary:
    overall_score: 0.87
    status: "PASS|FAIL|REVIEW"
    recommendation: "string"

  metrics:
    accuracy: 0.92
    consistency: 0.88
    robustness: 0.79
    format_compliance: 0.95

  test_results:
    total: 50
    passed: 43
    failed: 7
    pass_rate: 0.86

  failures:
    - id: "TC023"
      category: "edge_case"
      issue: "Description of failure"
      severity: "high|medium|low"
      input: "..."
      expected: "..."
      actual: "..."

  recommendations:
    - priority: high
      action: "Specific improvement"
    - priority: medium
      action: "Another improvement"

workflow:
  1_prepare:
    - Load prompt under test
    - Load test cases
    - Configure metrics

  2_execute:
    - Run all test cases
    - Record outputs
    - Measure latency

  3_score:
    - Compare against expected
    - Calculate metrics
    - Identify patterns

  4_analyze:
    - Categorize failures
    - Find root causes
    - Prioritize issues

  5_report:
    - Generate report
    - Make recommendations
    - Archive results

integrates_with:
  - prompt-optimization: Improvement feedback loop
  - all skills: Quality gate before deployment

automation:
  ci_cd: "Run on prompt changes"
  scheduled: "Weekly regression"
  triggered: "On demand"