E&C Forge
Designing a static GenUI MVP from zero: Intent → Result → Action · 2026

My Rol
AI Product Designer
IxD · Interaction Design
E&C Forge is our own startup, so there was no client brief to translate. I owned the product vision, the UX, the design system and the quality bar. Implementation was done with Claude Code, and I supervised it directly as QA-UX instead of handing specs to a team and hoping for fidelity.
What Makes This Project Different
This wasn't a redesign. It was a new product in a new paradigm, built from zero, where nobody had a validated mental model to copy. The product is also part of the argument: E&C Forge offers GenUI and accessibility-oriented design to mid-market US companies, so the site itself had to be the demo.
The Problem
Most company websites make visitors do the work: guess which menu item holds the answer, click through sections, assemble the picture themselves. For a studio selling a new way of building interfaces, that is the wrong first impression.
The bet behind E&C Forge is to prioritize user intent. Instead of navigating a fixed tree, the visitor says what they want, and the interface responds with a simple, relevant result and a clear next action. This first MVP is a static GenUI: a conversational prompt, guided chips and generated response cards inside a controlled layout, rather than a fully open generative canvas.
The risk was just as clear. A new paradigm has no conventions, and I had seen how products in this space, such as Google's Disco from Google's AI lab, can ship powerful ideas with little UX scaffolding. I wanted the opposite: an experimental concept on a disciplined UX foundation.

Process — From Structure to Interaction








UI: Light and Dark Modes, Responsive by MUI Guidelines










Design-to-Code: Figma MCP + Claude Code
Design, documentation and code were connected end to end: Figma → Figma MCP → Claude Code. That changed my role. Instead of writing a spec and waiting, I worked as QA-UX directly on the implementation, correcting spacing, states, behavior and copy with the agent in the same loop.
Some interaction details I had specified for years but rarely seen implemented, because they are slow and expensive with human teams. Tognazzini-style latency handling, with distinct feedback depending on how long the wait is, is the clearest example. With the agent, it could be implemented and verified in the same session.
Accessibility Built In
Screen reader support, full keyboard navigation, and contrast tuned for different visual conditions, including color blindness and low vision. These were implemented and checked by me; validation with assistive-technology users is part of the next round.
Validation — Where the Product Got Real
I believed the product was solid. The tests proved otherwise, and that is exactly why they were worth running.
Methods
Guerrilla tests
Tests with friends and professionals
Professional usability test: 5 participants, moderated, think-aloud, 4 fixed phases, 10 tasks
Synthesis: 47 atomic notes → affinity diagram (NN/g method) → 9 categories, each rated for frequency, severity (0–4) and Nielsen heuristic
Cross-check: formal heuristic scoring on top of the qualitative synthesis. Two independent methods converged on the same findings.


What worked
All 5 participants described the experience positively ("different", "competitive", "I'd never seen anything like it") and valued the conversational freedom. Nobody rejected the paradigm. That is the thing to preserve, not fix.
| # | Finding | Frequency | Severity |
|---|---|---|---|
| 1 | Minimal first impression read as “unfinished page” | 4/5 | Major |
| 2 | Blank prompt: not knowing what or how to ask (incl. language expectation) | 3/5 | Major |
| 3 | Unclear boundary between generated and static content | 3/5 | Major |
| 4 | History and edit functions not recognized | 3/5 | Major |
| 5 | Pricing: all 5 had to ask the prompt, no visible shortcut | 5/5 | Major |
| 6 | No state signal right after sending a message | 2/5 | Major |
| 7 | Contact CTA recognized inconsistently | Mixed | Minor |
| 8 | Agent translated “CEO” as “CTO” when describing the team | 1/5 | Catastrophic |
The last one is the finding I'm proudest of catching. It occurred once, but it isn't a perception problem; it's a content-reliability problem in the generation pipeline. I rated it by impact on brand credibility, not by frequency, and separated it from the visual-signal issues so it wouldn't be diluted into a "learning curve" story.
The research question, answered honestly
Is GenUI the problem, or is it missing interaction signals around a concept that works?
Six of the nine categories were missing signals (visibility of system status, recognition over recall, help, consistency), exactly what the hypothesis predicted. Finding #8 was a different kind of risk and needs engineering, not copy.
Iterations Driven by Research
• P0 — Facts about the company come from a structured source of truth, not free generation (CEO/CTO error)
• P0 — "Pricing" chip returning one prioritized answer with an indicative range, not a list of options
• P1 — Example chips + a framing line before the first interaction, so the page reads as intentionally generative
• P1 — Persistent history and edit icons with text labels and aria-label, never icon-only
• P1 — Visible waiting state between send and response (skeleton or indicator)
• P1 — One visual treatment for cards in chat and in static sections, with one source of truth per content type
• P2 — Real example placeholder in the prompt and a visible note that Spanish works
• P2 — Single recognizable contact CTA across chip, card and header
• Preserve — the open prompt. Fixes are scaffolding around it, never a forced wizard.
Design System — From MUI to a Proprietary System
Started from MUI's UI kit and extended it into a documented system with tokens, components, states, tone & voice and interaction patterns for both modes.
Pattern 01 — Latency & waiting states
Pattern 02 — Empty states (what it is + why + how to start)
Pattern 03 — Error handling with recovery paths
Pattern 04 — Generated vs static content signaling
Why These Patterns Matter Beyond Design
They are the shared language between me and the AI agent. Documented rules replaced repeated corrections.
Current Status and Next Steps
MVP live at ecforge.com. Business outcomes aren't measurable yet

