@alphabatcher: 25 render tools can cost ~10,000 tokens before the user asks anything Generative UI has 3 build patterns Controlled > p…
Summary
Discusses three build patterns for generative UI (controlled, declarative, open-ended) and provides a decision tree for choosing the right approach based on use case.
View Cached Full Text
Cached at: 06/04/26, 01:57 AM
25 render tools can cost ~10,000 tokens before the user asks anything
Generative UI has 3 build patterns
Controlled
pre-build React components register each one with
useComponentagent picks the component best for top 3-10 flows breaks when chart/table/card tools start sounding alike
Declarative
agent returns A2UI JSON frontend maps schema nodes to React components one function can drive many cards, tables, forms, widgets best for dashboards, search results, travel cards, reports breaks when
CATALOG_IDandcatalogIddo not match
Open-ended
agent writes raw HTML app renders it in sandboxed iframe best for one-shot visualizations breaks when sandbox flags block buttons or forms
Decision tree:
pixel-perfect mockup -> Controlled dozens of card types -> Declarative disposable visualization -> Open-ended
Audit your app:
count render tools if past 15, start wiring A2UI if using raw HTML, set iframe sandbox to allow scripts and forms only never allow-same-origin
Similar Articles
@Saboo_Shubham_: https://x.com/Saboo_Shubham_/status/2062220865643982875
The article explains three patterns of Generative UI (Controlled, Declarative, Open-ended) and how CopilotKit implements them via the AG-UI protocol for building dynamic agent-driven interfaces.
@tom_doerr: Generates UI with 67% fewer tokens than JSON https://github.com/thesysdev/openui…
OpenUI is an open-source framework for generative UI that uses a streaming-first language to produce UI with up to 67% fewer tokens than JSON, including a React runtime and chat interfaces.
@VraserX: This looks like a pretty big step for AI design. What stands out to me is that this is not just about generating pretty…
Genspark launches Genspark Design, an AI design tool powered by Claude Opus 4.7 that can create UI prototypes, posters, videos, HTML animations, and convert designs into code, aiming to be a full creative production tool.
Generative AI still feels stronger at making objects than letting people operate on them
The article discusses a common pattern in generative AI products: while initial creation has become easy, controlled revision and local editing of generated objects remain underdeveloped, limiting practical utility.
@heygurisingh: Holy shit... a startup just dropped a free open-source spec that generates UI components with 67% fewer tokens than JSO…
OpenUI is a new open-source specification that generates UI components with 67% fewer tokens than JSON, offering a model-agnostic, framework-neutral solution for LLM-based interface generation.