THE ENGINEER WHO BUILT CLAUDE CODE UNINSTALLED HIS IDE AND RUNS SEVERAL THOUSAND AGENTS OVERNIGHT FROM HIS PHONE. Here is the exact three-tier stack that makes it possible. Tier 1 is loops. Session-scoped. Daytime. One line runs a slash command on a fixed schedule inside an open session. The seven loops worth running in any session: /loop 5m /babysit # PR reviews, CI failures, conflicts /loop 30m /slack-feedback # mine Slack feedback into PRs /loop /post-merge-sweeper # sweep review comments after merges /loop 1h /pr-pruner # close stale PRs /loop 15m /triage-issues # classify and assign GitHub issues /loop 2h /claude-md-distiller # mine corrections into CLAUDE.md /loop 5m /deploy-watch # watch deploys, ping on regressions Slash commands live in .claude/commands as Markdown files checked into git. Build the workflow once. Loop it with one line. This is what the gap between using Claude Code and building with Claude Code actually looks like.(CyrilXBT)