Analyze classification errors to create targeted refinement instructions.

Task: {task_name}

Weak classes (need more training data):
{weak_classes}

Confusion patterns (true_label -> predicted_label):
{confusion_pairs}

Error samples:
{error_samples}

For each weak class, provide:
1. What distinguishes it from classes it's confused with
2. Characteristics to emphasize in new samples
3. Patterns to avoid (cause misclassification)

Return JSON:
{{
  "class_label": {{
    "confused_with": ["class1", "class2"],
    "differentiators": ["key difference 1", "key difference 2"],
    "emphasize": ["trait to highlight"],
    "avoid": ["pattern causing confusion"],
    "languages": {{"en": 0.5, "ar": 0.3}},
    "length": {{"min": 10, "max": 200}}
  }}
}}
