Net +283 Lines and a Passport

The key to pushing 29 PRs in a single day wasn't more automation, but a method of separating judgment from execution.

The work started from 30 tickets in Linear, and ran 4 concurrent Claude Code sessions across 8 waves. Each session first made a plan and then implemented it, and along the way a ship script was even built to link PRs, Linear tickets, and Threader sessions, reducing the friction of repetitive tasks. The result was a product reworked with +3,917 lines added, -3,634 lines deleted, and +283 net lines.

The most important judgment calls were left to humans. Vague tickets like "there's still a lot to look at on the agent page" were canceled at the planning stage, and plans that missed dependencies were not approved. Conversely, clear tasks were sent through in parallel.

This flow led to the same question in two completely different situations.

• Fast-growing teams worry about how to enforce quality as they scale.
• Teams running at full speed without a safety net worry about how to secure guarantees while maintaining speed.

The solution was standards-as-checks. Each PR had to pass 27 checks covering code conventions, security, mobile layout, accessibility, DB migrations, and more, and human preferences and standards were recorded as markdown inside the repo and executed as AI checks. Humans set the standards, the checks enforce those standards, and humans make the final call on what gets shipped.

The author also explains this in terms of "intervention rates." The metric that matters isn't how fast you shipped, but how often humans had to intervene. The better the checks are designed, the less often humans need to step in—and only at the precise points where it matters.

Renewing a passport works on the same principle. If you have the required documents, photo, and forms, you pass; if not, you're rejected. The system doesn't read your story—it only checks whether the requirements are met. The core point of this piece is that a codebase can be run the same way.