How to Make Claude Code Your AI Engineering Team

The video argues that we have entered a new agent era where the key to getting AI to do real work is giving it roles, process, and review—much like managing a human team—and that the bottleneck is no longer model intelligence but structure. The speaker demonstrates this with GStack, a platform built on "skills" that keep agents focused, such as an office-hours skill that interrogates a startup idea rather than blindly executing it, forcing founders to validate demand and reframe their product as a wedge strategy. The workflow evolves from planning to an adversarial review that automatically caught and fixed 16 issues, design variants that were scored, and auto-plan pipelines that run executive and engineering reviews. A standout feature is the Playwright CLI, which embeds a full browser into the tool, enabling autonomous testing and driving a custom QA system that catches real bugs. Ultimately, the message is that the barrier to building software has collapsed, leaving only the question of what you will build.

• ▶ 0:45 We are in a completely new agent era: getting agents to do real work means giving them roles, process, and review—just like a human team.
• ▶ 1:11 Gary has coded more in the past two months than in 2013, and rebuilt all of Posterous (originally 2 years, $10M, 10 engineers) using modern AI tools.
• ▶ 1:52 The real bottleneck is not model intelligence: out-of-the-box models wander and guess, producing plausible code that silently breaks—so GStack uses a “thin harness, fat skills” approach where structured skills like Office Hours keep agents on track.
• ▶ 3:14 GStack is built around “skills,” and the demo’s first skill — office hours — distills YC partners’ advising experience into a conversational coaching mode that interrogates the startup idea rather than just executing it.
• ▶ 5:28 The key turning point: the model asks, “What’s the strongest evidence that you have that someone actually wants this?” — which forces the founder to validate demand before building.
• ▶ 6:33 The model reframes the app as a wedge strategy: the 1099-finding feature is the hook, but the real business is lead generation for tax preparers, making monetization potentially 10x higher than a simple subscription.
• ▶ 10:14 The idea evolved into a browser automation flow: search Gmail for 1099s, ask which bank portals to add, log in, download PDFs, and email the CPA—skipping Google OAuth entirely.
• ▶ 13:02 An adversarial review automatically caught and fixed 16 issues across two rounds, improving the design doc from 6/10 to 8/10 before approval.
• ▶ 14:40 Three design variants were generated for the checklist dashboard; Option B (friendly progress) got a 5/5 and appeared to be the pick, while Option C was rejected for being overcomplicated.
• ▶ 16:30 Auto-plan runs the full CEO, engineering, design, and developer experience reviews automatically using the speaker's default recommendations, letting users skip extensive back-and-forth.
• ▶ 17:01 Cloud Code automatically builds the code upon plan approval, and the review command acts as a staff-level bug-catching service that catches issues missed during planning.
• ▶ 17:22 The Playwright CLI—embedding a full browser into the tool—is highlighted as a standout feature that enables autonomous browser-based testing and interaction.
• ▶ 18:10 Automating QA became a top priority after the agent handled planning, design, and coding, since manual QA was the least fun part.
• ▶ 18:48 Built SLQA and SL Browse by wrapping Playwright at the CLI level, enabling agents to drive real browsers, run regression tests, and assess real bugs.
• ▶ 21:06 GStack is available now, and the barrier to building software has collapsed—so the only question is what you'll build.

• ▶ 0:09 Introducing Gary, GStack, and the Thin Harness Problem (0:09 - 2:49) - - Gary introduces the agent era, GStack’s origin, and why successful agents need a thin harness with fat skills.
• ▶ 2:49 Live Demo: Office Hours for a 1099 Tax App (2:49 - 10:16) - - He demos Conductor/office hours on a 1099 tax app, including browser automation, bug-fixing, and three candidate approaches.
• ▶ 10:16 Refining the Approach and Choosing a Design (10:16 - 16:05) - - He refines approach B, runs adversarial review and design shotgun, then evaluates and locks in design variant B.
• ▶ 16:05 Cloud Code, Auto-Plan, and Post-Code Review (16:05 - 18:10) - - Covers Cloud Code adoption, auto-plan, Playwright CLI, parallel agents, and post-code review.
• ▶ 18:10 QA, Shipping, and Closing Pitch (18:10 - 21:44) - - Addresses QA bottlenecks via SLQA/SL Browse, walks through the ship tool, and closes with GStack availability.