A Mercari PM intern built mercari-pm-agent, a Claude Code Skill that automates the end-to-end PM workflow from problem discovery through PRD creation and UI mockups. The agent connects Notion, Slack, Looker, and Figma via MCP servers, querying them in parallel for data collection. Key design decisions include separating behavior definitions from reference templates to combat the 'Lost in the Middle' LLM problem, defining explicit constraints to prevent hallucinations (requiring PM confirmation before advancing steps), and using a dedicated evaluation skill to score output quality in a TDD-like loop. The post shares concrete lessons: constraints matter more than commands, prompt separation of concerns improves accuracy, and evaluation criteria should be defined before implementation.
