How a Dev Team Lead Redesigned Their Day with Claude Code

The morning routine of digging through Slack, Jira, and Confluence to prepare for the day was redesigned with Claude Code. At first, data was collected directly via MCP, but token cost and speed became problems, so the structure was eventually changed to separate collection into scripts, leaving the AI to handle only summarizing, judgment, and page generation.

As a result, the daily briefing was compressed into a single 15-minute page created by cross-referencing raw data from multiple systems. By breaking collection into four stages—collection → summary by person/task → briefing page generation → page upload—the results also became more stable, and the intermediate output, the summary DB, was reused for weekly documents, Q&A, and town hall preparation.

Important task discussions were conducted in a roundtable. Over the past month, 13 tasks and 19 ideas went through this loop, and operational policies, terminology, and decisions were recorded in files, accumulating as context for the next discussion.

In particular, for the shipping efficiency task, four perspectives—shipping operations, shipping product, delivery operations, and delivery product—were simulated in rotation to find connections that could easily be missed from a single perspective. While there is a limitation in that AI can only answer within the scope of what it knows, continuously accumulating context greatly improves the starting point for judgment.

The foundation underlying all of this is managed by a team-lead-role project. Each workflow's CLAUDE.md imports common rules, and at the end of each session, issues such as violations, recurrences, complaints, and environment, as well as the health of the rule repository, are checked, and improvement proposals are suggested.

Approved improvements are collected in each workflow's inbox.md and then promoted to common rules during the team lead session. As common rule distribution, the self-improvement loop, and the full propagation of improvements cycle through this process, the conclusion emerges that using tools well is not a matter of prompting skill but a matter of system design that raises the baseline capability.

Finally, an experiment has begun to extend this principle from the individual to the organization. Exploring where the bottlenecks lie, where the boundary between AI and humans should be, and how to build shared organizational context, the effort has entered a stage of transferring the structure that transformed one individual's day into the way the organization works.