Syntax - Tasty Web Development Treats 988: Cloudflare’s Next.js Slop Fork
80 snips
Mar 18, 2026 
Steve Faulkner, Cloudflare engineering director and creator of vinext, joins to unpack rebuilding Next.js with Vite in a week. They dig into AI coding workflows, markdown planning, agent browsers, and why tight feedback loops matter. It also gets into messy AI code quality, security testing, framework porting vs switching, and what AI-first languages and tooling might look like.
AI Snips
Chapters
Transcript
Episode notes
The Barbell Workflow Behind The Weekend Build
- Steve Faulkner let OpenCode run in long stretches overnight, then returned for short correction bursts during family time.
- OpenCode's session analysis showed peak token usage at 3 a.m. and a barbell pattern of two-to-four-minute sessions or one-to-two-hour deep dives.
Use A Minimal AI Stack With A Few Helpful Tools
- Keep your AI setup minimal and add only tools that clearly improve retrieval or execution.
- Steve Faulkner mostly used the OpenCode desktop app with no custom agents, then kept Context7, Exa Search, and Vercel's Agent Browser because they made results modestly but consistently better.
Feedback Loops Matter More Than Fancy Prompts
- Agent browsers work best when they can compare real UI behavior and record discoveries instead of rediscovering the same failures.
- Steve Faulkner had Agent Browser detect janky scroll behavior, then used discoveries.md to log ecosystem bugs and session compactions to preserve hard-won context.