No Figma. No Jira. No docs. How Gusto built a new product line with Claude Code | Eddie Kim (CTO)

What It Covers

Gusto CTO Eddie Kim describes how four engineers and one designer built Gusto Cofounder, a new AI-powered product line, in ten weeks with zero meetings, no Figma files, no Jira board, no tech specs, and no dedicated product manager — using only Claude Code and a permanent Zoom room.

Key Questions Answered

• •Zero-documentation development: Eliminate tech specs, Figmas, Jira boards, and stand-ups for zero-to-one builds. Gusto's five-person team operated from a single whiteboard photo for ten weeks. The only structured process was a 24/7 open Zoom room where engineers dropped in to share screens, discuss features, and conduct code reviews in real time.
• •Trash-can engineering method: Treat code as disposable by opening real, human-ready pull requests and then closing them if the feature fails a group discussion. With AI-assisted coding, rebuilding from a v2 branch from scratch is faster than refactoring. This removes emotional attachment to code and accelerates product-shape decisions without a product manager.
• •Minimal agent stack: Build AI agent loops with two tools only — Cloudflare Workers for the agent runtime and Vercel AI SDK for model flexibility. No third-party memory or planning harnesses are needed. Memory is simply a tool that writes to a database column. This stack removes complexity and lets small teams ship production-grade agents quickly.
• •Non-engineer PR throughput: Gusto's designer, Katie, ranked in the 94th percentile for PR throughput across the entire 1,000-person R&D organization. She achieved this by pairing with engineers who reviewed her Claude-generated code, gave prompting feedback, and helped her develop taste for code quality — demonstrating that design-to-production contribution is replicable with deliberate mentorship.
• •Eval-driven bug fixing: When addressing customer-reported issues, write a failing eval first to reproduce the problem, then prompt Claude Code with a GitHub issue link to generate a fix, and confirm the solution by verifying the eval passes without breaking the existing eval suite. This replaces traditional debugging and produces a reviewable pull request automatically.