Building Software That People Love

What It Covers

Wesley Yu, VP of Engineering at MetaLab — the design studio behind Slack, Uber, and Instacart's early product DNA — explains how small cross-functional teams build delightful digital products by combining boring tech stack choices, iterative ugly-app development, and tight design-engineering collaboration to ship v1 products for ambitious early-stage companies.

Key Questions Answered

• •Tech Stack Selection: Agencies should build a decision tree prioritizing client maintainability over engineering preference. First ask what technologies the client's team already knows, then what they can hire for in their geography — European clients often require Java/Spring Boot due to university pipelines — and finally what energizes their technical leads, since a stack choice is simultaneously a hiring plan and a 2AM incident response plan.
• •Ugly App Development: Build end-to-end flows immediately, even if intermediate screens render raw JSON with a single button. This iterative approach — moving from JSON to wireframes to mid-fidelity to high-fidelity — lets clients see the full experience early, understand feature complexity in real time, and reorder screens before polish is applied, preventing costly late-stage structural changes.
• •Risk Isolation via Sidecars: When a project requires an unproven technology, contain it in an isolated sidecar service rather than integrating it into the core monolith. For example, high-performance real-time WebSocket behavior can be handled by an Elixir/Phoenix sidecar, keeping the modular monolith stable while limiting the blast radius if the experimental component needs replacement.
• •Warm Handoffs Over Documentation: Cold handoffs — leaving comments, video walkthroughs, and README files — are the worst-case scenario for transferring a codebase. Instead, run three to four sprints with the incoming maintainer embedded in PR reviews and technical decisions before full transfer. Advise clients to make these hires early so the incoming engineer's preferences shape decisions while changes are still low-cost.
• •AI Tooling on the Frontier: Building bespoke solutions on emerging AI infrastructure wastes significant engineering time. MetaLab spent heavily on custom LLM eval frameworks and custom MCPs for Notion and Figma, only to discard them when LangFuse and official vendor MCPs launched. The strategic question is identifying which frontier investments to make versus which to wait for the industry to standardize before adopting.