Back office ops · Production

Mercari PM intern builds mercari-pm-agent automating end-to-end PM workflow via Claude Code Skills and MCP

The problem

Mercari PMs needed to gather information from Notion, Slack, Looker, and Figma before every product decision, and the time spent moving across tools and organizing which data mattered was taking PMs away from deeper thinking and stakeholder conversations.

First attempt

Consolidating all agent behavior definitions into a single SKILL.md file produced poor output accuracy due to the LLM 'Lost in the Middle' problem, where models fail to attend to relevant information in long contexts.

Workflow diagram · grounded in source
1
PM describes business problem
trigger
“When a PM describes a business problem on the product in natural language, the following steps proceed automatically.”
2
MCP data collection
integration
“connected Notion, Slack, Looker, and Figma via MCP (Model Context Protocol)”
3
PM confirmation gate
human_review
“I prohibited the agent from automatically progressing to the next step without PM confirmation. You are NOT allowed to infer completeness. Only explicit confirmation from the PM allows progression.”
4
PRD creation
output
“Integrate all of the above into a PRD (Product Requirements Document)”
5
UI mockups generation
output
“problem discovery → data gathering → PRD creation → UI mockups”
6
Evaluation skill feedback loop
feedback_loop
“I created a dedicated evaluation skill (skill-creator-max). It sends test cases to mercari-pm-agent and returns scores for output quality. Through iterative improvements using this score, I obtained the insight mentioned earlier”
Reported outcome

Separating SKILL.md from reference files produced a clear improvement in evaluation score, and the agent can now run the full PM workflow end-to-end within a single session.

Reported metrics
Output accuracy after file separationclear improvement in score
Data collection waiting timereduces waiting time compared to sequential access
Reported stack
Claude CodeNotion MCPSlack MCPSocratesFigma MCPBigQueryLookerNotionSlackFigma
Source
https://engineering.mercari.com/en/blog/entry/20260427-mercari-pm-agent-design-automating-the-pm-workflow-with-claude-code-skills-and-mcp/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Separating SKILL.md from reference files produced a clear improvement in evaluation score, and the agent can now run the full PM workflow end-to-end within a single session.

What tools did this team use?

Claude Code, Notion MCP, Slack MCP, Socrates, Figma MCP, BigQuery, Looker, Notion, Slack, Figma.

What results were reported?

Output accuracy after file separation: clear improvement in score; Data collection waiting time: reduces waiting time compared to sequential access (source-reported, not independently verified).

What failed first in this deployment?

Consolidating all agent behavior definitions into a single SKILL.md file produced poor output accuracy due to the LLM 'Lost in the Middle' problem, where models fail to attend to relevant information in long contexts.

How is this back office ops AI workflow structured?

PM describes business problem → MCP data collection → PM confirmation gate → PRD creation → UI mockups generation → Evaluation skill feedback loop.

Related back office ops cases