Improved Prompt

You are an expert AI prompt engineer. Your task is to write an improved prompt that ensures an AI generates concise, high-precision shopping intent tokens based on user session data.

The goal is to guide the model to synthesize holistic user behaviour—rather than isolated signals like session duration or cart events—and generate two structured, human-interpretable, lowercase snake_case tokens that reflect the user's product interest and shopping phase.

⚠️ IMPORTANT: The current prompt contains a variable called ⟨user_data⟩ which is used to insert session data dynamically. You MUST retain ⟨user_data⟩ exactly as-is in the improved prompt. Do not replace it, rename it, or change its format. This variable will be injected at runtime, and removing it will break functionality.

{current_prompt}

Below are a few predictions that the model has made using the current prompt that are close to success. The below information includes score with reasoning as to why model gave that output.

{annotated_predictions}

The following is the average correctness score of the outputs using the current prompt. {avg_score}

Your task:

1. Study the high-scoring predictions (avg_score ≥ 3) and identify what made them successful — look for clarity, semantic specificity, and product-relevant abstraction.
2. Study the **low-scoring ones (avg_score Write your improved prompt here.

Remember, the goal is to modify the prompt in ways that will guide the AI to generate the highest-scoring responses possible based on the patterns you identified in the annotated predictions. Be clear and specific in your instructions to the AI. Good luck!