How to Make Claude Code Your AI Engineering Team

Tan argues that agents do useful work when they have roles, process, and review. He built GStack as a thin harness with specialist skills, saying the main problem is often missing context and unchecked guesses rather than model intelligence.

Tan demonstrates GStack’s Office Hours skill by exploring an app that finds tax documents such as 1099-INT forms. Its questions probe customer evidence, the pain point, and existing alternatives, then help reframe document collection as a possible path to tax preparation referrals.

Office Hours produces three approaches, from a simple Gmail checklist to browser automation paired with a CPA marketplace. Tan favors the broader browser-based concept, then describes an adversarial review that found and fixed 16 issues across two rounds. Design Shotgun generates three dashboard options, and Tan selects a friendlier, card-based layout.

Tan outlines a workflow that includes planning, CEO and engineering review, coding, and a staff-level code review. He says GStack’s Playwright and Chromium tools let agents inspect pages, reproduce browser bugs, and run regression checks. He uses multiple parallel sessions and worktrees to handle separate projects and fixes.

Tan describes a final ship step for checking whether a pull request is ready to merge. He says he runs 10 to 15 Claude Code sessions and can process 10 to 50 pull requests in a day, depending on meetings. He also says he reviews community contributions carefully because he is concerned about supply chain attacks.