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

The podcast details a project that achieved rapid product development with a minimal setup, avoiding traditional planning tools like Figma, JIRA, or tech specs. The core system utilized Cloudflare Workers for the agent loop and Vercel AI SDK for model switching and tool integration, emphasizing simplicity over complexity. A small team of five (four engineers and a designer) built the product in 10 weeks using an agile, iterative approach, including the "trash can method" of discarding and rebuilding code as needed. Decision-making relied on collaborative discussions and immediate code reviews, with low-cost experimentation enabling flexibility. The project highlights a shift toward speed, adaptability, and AI-driven development, contrasting with traditional planning-heavy methods. This approach underscores potential for scaling similar practices in larger organizations by leveraging AI to accelerate R&D and reduce bottlenecks.

The development process emphasized unstructured collaboration, such as a 24/7 Zoom room for real-time input, and minimal formal documentation, relying instead on a single whiteboard wireframe. A non-engineer (a designer) contributed significantly, achieving high PR throughput through technical curiosity and mentorship. The team prioritized shipping early and iterating, using fake UI experiences in production and gradually replacing them with functional backend logic. Technical execution focused on avoiding over-engineering, with tools like AISDK simplifying agent loops. The project also underscores the value of unstructured time for innovation and the need for inclusive, cross-functional workflows, where non-engineers actively contribute to code and design. Overall, the approach challenges traditional development norms by promoting speed, flexibility, and the efficient use of AI tools to streamline product creation.

• What if you built an AI agent using only Cloudflare Workers and the Vercel AI SDK, avoiding external dependencies?

  • Move: Use Cloudflare Workers for the agent loop and Vercel AI SDK for model switching, skipping traditional tooling like Figma or JIRA.
  • Why Now?: These tools are lightweight, cloud-native, and require minimal setupideal for solo developers needing speed and flexibility.
  • Expected Upside: Rapid deployment of a functional AI agent with minimal infrastructure overhead, leveraging modern tooling to reduce bottlenecks.

• What if you adopted the "trash can method" to rebuild features from scratch if they no longer align with your priorities?

  • Move: Start a "V2" branch for critical features, delete non-essential code, and rebuild iteratively using pull requests for immediate feedback.
  • Why Now?: The low cost of experimentation (e.g., no team dependency) allows solo operators to pivot quickly without long-term commitment.
  • Expected Upside: A leaner, more adaptable codebase that evolves with user feedback, avoiding technical debt from outdated features.

• What if you used a "perma-zoom" (always-on collaboration tool) for real-time feedback on your code, even if its incomplete?

  • Move: Host a 24/7 Zoom session (or equivalent) to share WIP code with peers, using informal reviews to iterate on design and functionality.
  • Why Now?: Collaborative, asynchronous feedback loops accelerate learning and reduce the need for formal documentation or meetings.
  • Expected Upside: Faster problem-solving, reduced isolation, and alignment with user needs through continuous input, even for small-scale projects.

• Leverage minimal tooling stacks: Use Cloudflare Workers for agent loops and Vercel AI SDK for model switching/tool integration to avoid overcomplicating your tech stack with meetings, Figma, or JIRA.
• Adopt the trash can method: Build features iteratively, review them immediately, and delete or rebuild them if they dont align with priorities, reducing emotional burden and enabling flexibility.
• Collaborate without formal documentation: Use a shared whiteboard and 24/7 Zoom room for real-time idea sharing and code reviews, minimizing reliance on PRDs, Figma, or tech specs.
• Involve non-engineers in technical workflows: Assign non-software engineers (e.g., designers) to contribute code, participate in PR reviews, and iterate on UI/UX, ensuring cross-functional collaboration.
• Ship early with fake UI/production experiments: Launch MVPs with fake UI or placeholder logic using feature flags, then gradually replace them with real agent logic while maintaining a consistent frontend.