AI-Assisted UX Discovery with a Coding Agent — The Rejection Loop Method

Most "AI in design" advice optimizes the wrong thing. It celebrates generation — faster mockups, more variants — when the real bottleneck in UX discovery is rarely generation. It is grounding (does this idea survive contact with the product's data?) and rigor (is this idea even testable?). Without those, a discovery sprint produces opinionated comps that get rebuilt three times by engineering and ship with placeholder data, or worse, never ship at all because no one named a metric.

A 7-step loop, run with an AI coding agent as the workshop partner: (1) frame from a quantitative scorecard plus business and SEO non-negotiables, (2) force every opportunity into a structured hypothesis-card with a primary metric, (3) ground each design region in the actual product schema, (4) generate mid- to high-fidelity annotated concepts conditioned on a real design system, (5) run a cross-model critique pass that returns "top 5 to change, ranked by impact," (6) capture as a shareable summary plus image, (7) exit with testable next steps and named owners.

Designs that survive contact with the real data model — in minutes, not days. The agent reads your GraphQL or REST schema and maps every region of every concept to a real field. Fantasy designs get killed before they reach Figma. The second win is the cross-model critique loop: a vision-capable model grades the image-gen model's output against the hypothesis and surfaces structural problems the human would otherwise have to catch alone. The human stays focused on rejection and direction, which is the work that actually moves the discovery forward.

The worked example used Pi as the coding agent and gpt-image-2 as the visual generator. The critique pass should be run with a different model family — Claude Sonnet or Gemini Pro Vision both work — to avoid in-family bias. The method is tool-agnostic: any coding agent with file-read, schema-read, and image-generation tool access can run it. Three prerequisites are load-bearing: a quantitative scorecard, a design-system reference (tokens plus 5–10 component screenshots), and read access to the product's data model.

This is not a substitute for user testing — quantitative usability validation still requires real users. It will not do brand or visual-identity work; image generation does not reach that fidelity. If the agent cannot access the product's schema, the schema-grounding step degrades to a written exercise and the method loses much of its rigor. Image generation without a design-system reference lands in uncanny valley and gets rejected on aesthetics before the structural ideas can be evaluated. And solo mode requires unusual discipline — the Rejection Loop only works if the human exercises rejection.