PMs Writing Prompts and Developers Explaining AI

Following its KakaoTalk ChatGPT integration in March 2026, Olive Young officially launched a dedicated ChatGPT plugin in August. This allows users to enjoy product recommendations reflecting skin type, ingredients, and discount status, as well as a search experience based on actual buyer reviews. As of July, monthly visitors to the Olive Young online mall via ChatGPT had increased by more than 30% compared to January.

The 'Hint Game' of Communicating with LLMs

Technically, implementing the MCP (Model Context Protocol) itself was a simple task requiring only compliance with the specifications of existing search services. However, in actual production environments, issues arose where the LLM failed to accurately grasp user intent. This situation was similar to a 'charades' game where the LLM must infer the correct answer based solely on hints from the Tool Description.

In particular, cases where the minPrice and maxPrice parameters were misinterpreted were frequent in price-range searches. For example, when a user entered '20 million won range', the LLM incorrectly interpreted it as '20,000 won range', or products in unexpected price ranges appeared under the '15000 won' condition.

Contrasting Solution Approaches by PMs and Developers

To address these issues, PMs and developers attempted different strategies.

• PM's Approach (Providing Few-Shot Examples): Based on the insight that LLMs judge better through concrete examples than abstract rules, various utterance types such as '20,000 won range', 'under 30,000 won', and '15000 won', along with their correct parameter values, were added to the Description as Few-Shot examples.
• DEV's Approach (Explaining Mathematical Rules): Judging that listing all possible utterance types as examples lacked scalability, they instead specified mathematical rules for calculating 'n ten-thousand won range' (e.g., 20,000–29,999 won for the 20,000 won range) to guide the LLM to calculate ranges using generalized logic.

This case demonstrates how important it is in AI service development to go beyond simple feature implementation and understand the reasoning characteristics of LLMs, meticulously refining prompts and parameter descriptions.