Skip to content

Just-In-Time Site Rendering

An experiment in personalized, adaptive web experiences — generating entire web pages at request time from site-owner facts, design systems, and a knowledge base.

workflowspage-generationcontent-authoring

Heads up: the audio on this recording is rough.

In late 2024 — before Design.md, Skills, Rules, agentic workflows, and AI harnesses were common vocabulary — this experiment explored personalized, adaptive web experiences and generating whole websites at request time.

The core idea

Page copy is derived from site-owner-supplied content to avoid hallucination. There's no need to pre-write articles or localize content up front — you just equip the site with the facts relevant to the business and have AI generate the pages as they're requested. Brand and design systems are provided as artifacts ahead of time, along with media and icons, which keeps a consistent voice and tone throughout the site and applies guardrails where needed.

Today's (yesterday's?) model latency is too slow to run this live in production, but a cache-warming agent can build the site, run evals over it, and push to production. A second SEO/AEO agent then evolves the content once it's live — letting the site owner focus on the business instead of content marketing.

How it works

Agent workflows run on each request. The page-rendering service receives the requested URL, locale, and domain, then identifies the page type from the URL and retrieves the relevant site artifacts — the design system and knowledge base the owners maintain. The workflow then:

  1. builds a wireframe from the components in the design system,
  2. produces copy from the knowledge-base articles, in the required voice and tone, and in the requested language for the locale.

In future it could go further — tailoring pages to the referrer. Arrived from Twitter? Maybe a short-form fact sheet. Came in on a book-review backlink? A long read might suit better.