One in three Vercel deploys this week was kicked off by a coding agent — up roughly tenfold in six months. This issue surveys what's moving in agentic web design, with a side of risograph.
The prompt-to-app market has narrowed to Lovable, Bolt.new, and Vercel's v0. The most consequential update this cycle is v0's February 2026 release, which pushed it past prompt-to-UI and into a real dev environment — Git integration, an in-browser code editor, database connectivity, and agentic planning workflows. As of March it's being called the strongest tool for production-quality React.
Lovable 2.0 is the more horizontal play: Lovable Cloud (built-in auth + persistence), real-time multiplayer for up to 20, an agentic mode for multi-step autonomous edits, AI connectors (Perplexity, ElevenLabs, Firecrawl, Miro), visual CSS editing, themes, analytics, and domain purchasing — closer to a "Webflow + Supabase + agent" hybrid than a pure code generator.
Bolt.new, meanwhile, has leaned into framework breadth where v0 is React-only — Next.js, Astro, Remix, Svelte, and Vue all in scope, with StackBlitz's WebContainer running everything in-browser. The three tools occupy distinct positions now: v0 for React production, Lovable for end-to-end app builds with backend wiring, Bolt for stack flexibility.
Component prompting beats whole-site prompting. The advice converging across builder blogs this spring: don't ask for a full site in one shot. Ask for one section at a time ("a high-conversion hero for a vegan bakery"), refine it, then move on. This pairs naturally with v0's pattern of generating three visual variations of one component.
The PRD-shaped prompt. Structure your prompt as a mini Product Requirements Document — Identity + Tech stack + Vibe + Core features. Pair that with explicit edge cases, security constraints, and testing expectations and output quality jumps measurably.
Context layering. Vibe-coding writers are converging on this term — stacking architecture notes, recent commits, API contracts, and user stories into the prompt so the model starts with the same mental map you'd hand a new hire. Without it: boilerplate. With it: code that fits the repo.
Plan-first workflows. If a feature has 20 decision points and the agent gets each right 80% of the time, the odds of a clean run are about 1%. Stack Overflow and Zencoder both made this point in March: planning collapses those 20 decisions into a reviewed spec.
An April piece in Design Systems Collective lays out a workflow gaining real traction: use Claude Design for visual exploration, then hand a fully-specified design system (color, type, components, spacing) over to Claude Code for production. Designers say the two-tool split is what makes the difference — mixing them is where beginners burn hours.
The setup that pays off, per Katherine Yeh's designer's guide: drop actual brand assets — logo, brand guidelines doc, reference imagery — into a brand_assets folder so the agent stops asking the same context questions on every session.
Related and timely: Julian Oczkowski's March piece on 7 Claude Code design skills that auto-activate when the conversation matches what they're built for, removing the "remember to invoke the right tool" tax.
A clean baseline that's been making the rounds — useful as a teardown of where prompts begin:
This is the floor of acceptable. Layering brand voice, tech-stack constraints, accessibility requirements, and a competitor URL turns it into a prompt that produces something you'd actually ship.
Three of the four sections of this issue point in the same direction. Agentic web tooling is consolidating around a small number of mature platforms, and the practitioners using them well are the ones treating prompts as specifications rather than wishes — PRD-shaped, context-layered, plan-first.
A year ago the conversation was about whether AI could ship a website. This week it's about how to brief the agent that's already shipping it. The skill is moving from generation to orchestration, and the studios pulling ahead are the ones building reusable design systems and brand-asset folders the agent can lean on — not the ones writing better one-shot prompts.