Margo Sultenfuss

← / CASE STUDIES

2025–2026 · Agent.ai

Generative UI Content Schema

Systems Design0-to-1

One content system governing what 24 live agents show, how they mark uncertainty, and how they handle thin data.

Role
Lead Product Designer, Agentic AI
Scope
24 live premium agents
Outcome
Validated by blind test; planned for platform-wide adoption
FIG. 01The missing layer, owned by content architecture.

24

Live premium agents

3

Parts: registry, rules, generator

1

Blind run, zero prior context

/ THE PROBLEM

Visual consistency was high, but output content drifted across every agent. No one owned the layer that governs it, so engineers were making product decisions in code while trying to ship.

Before → After

FIG. 02The same content type, rendered three ways by three agents, then resolved under one content system.

/ WHAT I BUILT

A content system in three parts: a registry of content types and the structural primitives that arrange them, the platform-wide behavioral rules every agent inherits, and a generator that turns a PRD into a reviewable output spec.

FIG. 03Mappings from content entities.

/ HOW I PROVED IT

A blind run: a fresh model, no history, given only the system and a new agent's PRD. What came back was consistent with everything else in the system, and it surfaced edge cases my own audit had missed, which went back into the platform rules.

/ Spec: financial-metric-card.spec.md

// Copy constraintmax_words: 20 - headline truncated at word limit
// Signal thresholduncertainty_threshold: < 0.70 → badge: "Low Signal"
// Null safetynull_fallback: "Data unavailable"

/ Rendered output (React)

AAria

Q3 Analyst Targets

-3.2%EPS vs. consensus
⚠ Low Signal

Analyst Note: data unavailable

/ The Long Version

Read the full case study

← / CASE STUDIES