Website Engine
Awwwards-level HTML direction
Constraints
- ■ Route briefs, section packs, and compiled prompts must remain connected to the generated Astro/static page artifact.
- ■ The pipeline needs browser review and scoring before a generated layout can be treated as production-ready.
- ■ Every generated section has to keep its purpose, proof role, and acceptance criteria traceable.
- ■ Regeneration must be possible without losing the original strategy, structure, or content contract.
Fragment Database
Variability Engine (XML Repository)
Deterministic Assembler
The Website Engine core. A Python script compiles the final prompt from dozens of XML fragments, resolves dependencies such as niche_name and injects content packs before the request is sent to the LLM.
Track 1: Single-Page
generate-v7.md
Build Context
Loads ux-psychology-rules.xml (Gestalt, Hick's Law), marketing-widgets, Style DNA, and reference-implementations before generation.
Create HTML Shell
Generates the base shell with 3-tier design tokens, Swiper initialization, and Lucide icons.
Section Assembly
Builds UI sections iteratively, applies the Von Restorff effect, and validates FAQ, mobile menu, and lightbox logic.
Quality Gates
Runs code_analyzer.py and auto_fix.py, then checks the output with score_bundle.py.
Visual QA
Checks Responsive Grade B+ (≥75), Structural Score ≥70, and
Final Report
Summarizes applied UX rules, auto-fix results, final scoring, and writes the run into generation_log.md.
Track 2: Multi-Page
generate-multipage-v2.md
Assemble Context
Runs assemble_multipage_prompt.py, prepares dynamic content packs, and builds _shared_context.md.
Generate Landing
Renders index.html and extracts , , and for reuse across the site.
Inner Pages
Runs the --pages routing loop, injects content packs such as team and FAQ data, and generates unique SEO metadata.
Bundle QA
Runs auto_fix.py in bundle mode, validates links with validate_bundle.py, and applies unified bundle scoring.
Final Report
Reports generated pages, content-pack integration status, and the final unified score.
Generation gallery
Beauty Salon
style: editorial_magazine
Fitness Club
style: cyber_hud
Handyman
style: industrial_tech
Interior Design
style: organic_natural
Law Firm
style: luxury_dark
Medical Clinic
style: soft_glassmorphism
Real Estate
style: high_end_corporate
Restaurant Cafe
style: organic_natural
Synthesis of technology and meaning
Website Engine combines four key product dimensions, generating not just markup but a launch-ready commercial asset.
Reliable foundation
Predictable HTML structure without AI drift. Clean responsive code and stable layout out of the box.
Visual DNA
Deep rebuild of design tokens: accent colors, typography, and distinctive micro-interactions.
Meaningful copy
No Lorem Ipsum. Smart copywriting and relevant imagery selected for the specific niche.
Conversion
Audience analysis and positioning built in. A commercial asset, not just a website.
Engine evolution through 7 iterations
From deep research of top design systems to autonomous pipelines for production-ready interface generation.
Research and source analysis
Mapped the source brief, domain signals, design constraints, and route goals before any generation step.
Prompt contract assembly
Assembled XML fragments, page briefs, section rules, and proof requirements into a reusable generation prompt.
Initial HTML generation
Generated the first complete HTML artifact with explicit layout, copy, component boundaries, and route intent.
Scaling and modularity
Moving to multi-page generation required a modular extension architecture.
Automatic expansion
Libraries expand automatically when new XML files are added to the corresponding folders. The engine picks up new niches, styles, and presets without changing assembler code.
Design and copy iteration
Separated visual polish, marketing copy, responsive fixes, and structural cleanup into deliberate review passes.
Quality scoring experiments
Validated output with scoring scripts, browser inspection, hierarchy checks, and acceptance thresholds.
Agent-ready route harness
Adapted the same route contract so future pages can be regenerated with context, gates, and artifact proof.
What this proves
The Truth
The case proves that AI-first delivery works when it is constrained by architecture, content models and verification gates.
Methodology
The method combines Astro-style static delivery discipline, route-level artifacts, browser review, and scoring gates instead of relying on a single generative prompt.
Use the same approach for production systems that need AI acceleration without losing control.
Related materials
Need a website engine that turns briefs into controlled production pages?
I can help design a deterministic AI-first pipeline for landing pages, case pages, content systems, and QA gates without turning delivery into guesswork.