# AI-Native Catalog Integration at Mercari with LLMs and a Prompt Generation Agent

DevFeed: [AI-Native Catalog Integration at Mercari with LLMs and a Prompt Generation Agent](<https://devfeed.tech/articles/ai-native-catalog-integration-at-mercari-with-llms-and-a-prompt-generation-agent-127380.md>)

Original publisher: [Read original article](<https://engineering.mercari.com/en/blog/entry/20261110-ai-native-catalog-integration-at-mercari-with-llms-and-a-prompt-generation-agent/>)

Author: DoiYutaro

Published: 2026-11-10T12:43:20Z

Content type: article

Language: en

Sources: [Mercari Engineering](<https://devfeed.tech/sources/mercari-engineering.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [AI Integration Strategies](<https://devfeed.tech/topics/ai-integration-strategies.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>), [Agent Skills](<https://devfeed.tech/topics/agent-skills.md>), [Claude Code](<https://devfeed.tech/topics/claude-code.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-native](<https://devfeed.tech/tags/ai-native.md>), [catalog](<https://devfeed.tech/tags/catalog.md>), [generation](<https://devfeed.tech/tags/generation.md>), [integration](<https://devfeed.tech/tags/integration.md>), [llms](<https://devfeed.tech/tags/llms.md>), [prompt](<https://devfeed.tech/tags/prompt.md>)

## AI overview

Mercari replaced Go rules for converting external catalog data with natural-language prompts executed by an LLM, and built an agent to investigate provider data, create prompts and schemas, and refine them against sample records. This lets Catalog Specialists adjust conversion rules directly and reduces engineering handoffs. Production safeguards include schema-constrained JSON outputs, code for checks where variance could harm data integrity, and a review-and-correction loop with dry runs. In a Pokémon-card offline evaluation, tuned prompts on gpt-5.4-mini reached 85.2% all-field accuracy across 15 fields, versus 69.6% for the rule-based method; few-shot-only prompting scored 66.2%. The more capable gpt-5.6 reached 98.7% when tuned, showing model differences remain. After rollout, Mercari reports integrating six categories at roughly the same headcount, with over 40 times as many records and about 55% less effort per new category. LLM fees are a new per-record cost. Periodic human review and production releases for prompt fixes remain manual bottlenecks.

## Source excerpt

Introduction Hello! We are arr0w (ML) and utdoi (Backend), Product Engineers at Mercari. Our team, the Product Data Team, works on integrating product data held by external businesses to improve the coverage and quality of our product catalog (data that consolidates information for each product). External product data can't be used as-is; it has to [...]