Type: url
Original: https://techblog.zozo.com/entry/ai-development-two-commands
Status: generated
Summary
ZOZO’s core-systems division standardized AI-driven development across the organization into two Claude Code commands, /dev-init and /dev-resume: the first builds a Confluence design document and a local progress table from a Jira ticket after parallel sub-agent investigation of the codebase, and the second reads that table against git status/git diff to re-establish where work stands. The post argues that what to delegate should be decided by the direction of information transformation — same-level and concrete-to-abstract conversions go to AI, while abstract-to-concrete judgment such as planning, requirement shaping and prioritization stays with humans. For quality-critical work it pairs Claude Code with Codex in a three-round critical dialogue in which Codex independently investigates the repository and criticizes, Claude Code adjudicates each finding as accepted, rejected or deferred, and Codex re-criticizes the rejections.