跳到正文
cccskills
免费在 GitHub 公开

ephemeral-generative-ui-architect

Expert guide for Generative UI architectures, ephemeral dynamic interfaces, and real-time streaming of React Server Components (RSC) driven by Gemini 4 Pro / Panduan ahli arsitektur Generative UI, antarmuka dinamis efemeral, dan streaming RSC real-time yang didorong oleh Gemini 4 Pro.

中文简介正在准备中。当前显示原文说明。

集成工具VercelGemini
查看安装方法

包含的文件(1)

  • SKILL.md6.7 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

Ephemeral Generative UI Architect (2026 Edition)

English | Bahasa Indonesia


<a name="english"></a>

English

Orchestration & Integration

Connects and orchestrates with vercel-ai-sdk-expert, senior-frontend, ui-ux-pro-max, design-system-architect, gemini-agent-booster, and multi-agent-orchestration.

Description

Standard UIs are static and pre-compiled. In the Gemini 4 Pro era, interfaces are ephemeral: generated in real-time based on the user's immediate intent, and destroyed when no longer needed. This skill orchestrates Generative UI systems using React Server Components (RSC), Tool-Driven UI injection, and Vercel AI SDK, allowing LLMs to directly stream fully functional, interactive components (charts, forms, dashboards) instead of markdown text.

Trigger Conditions

  • The user requests an interface that adapts entirely to the data context (e.g., "Show me a dashboard of my spending" -> AI generates a custom chart component on the fly).
  • Designing AI-native chat interfaces that render interactive widgets (booking forms, map components) instead of plain text responses.
  • Integrating Vercel AI SDK streamUI or tool-call-to-component mappings.

Core Architecture (2026 Standard)

  1. Tool-to-UI Mapping: The AI decides to call a tool (e.g., getWeather). The server intercepts this tool call and streams a <WeatherWidget /> React Server Component directly to the client.
  2. Zero-Client-State (Ephemeral): The UI state exists only within the LLM conversation history. If the chat resets, the UI resets.
  3. Optimistic Rendering: The client immediately renders a loading skeleton (e.g., <WeatherSkeleton />) the millisecond the LLM begins emitting the tool call token, replaced by the final RSC once execution completes.

Sovereign Visual Hierarchy in Ephemeral UI

Even when rendered dynamically on-the-fly, AI-generated components MUST strictly observe Sovereign Visual Hierarchy:

  1. Container Grounding: Wrap streamed widgets in Level 1 Surface containers (bg-white dark:bg-neutral-900 border border-neutral-200/80 dark:border-neutral-800 rounded-xl shadow-xs).
  2. Typographic Rhythm: Eyebrow badge (text-[11px] font-semibold uppercase tracking-wider text-neutral-500) -> Title (text-lg font-semibold tracking-tight text-neutral-900 dark:text-neutral-50) -> Body (text-sm text-neutral-600 dark:text-neutral-300).
  3. Tabular Numerals: Enforce tabular-nums on all streaming counters, prices, metrics, and telemetry stats.
  4. Single Primary CTA: Exactly one primary action per streamed card (e.g., "Confirm Booking"); secondary options use subtle ghost styling.

Implementation Recipe (Vercel AI SDK + React 19)

import { streamUI } from 'ai/rsc';
import { google } from '@ai-sdk/google';
import { z } from 'zod';
import { FlightTracker } from '@/components/flight-tracker';
import { Skeleton } from '@/components/skeleton';

export async function submitMessage(userInput: string) {
  'use server';

  const result = await streamUI({
    model: google('gemini-4-pro'),
    prompt: userInput,
    text: ({ content, done }) => {
      return <div>{content}</div>;
    },
    tools: {
      trackFlight: {
        description: 'Track a flight status in real-time',
        parameters: z.object({ flightNumber: z.string() }),
        generate: async function* ({ flightNumber }) {
          yield <Skeleton className="h-40 w-full" />; // Optimistic UI
          const data = await fetchFlightData(flightNumber); // Server-side execution
          return <FlightTracker data={data} />; // Final RSC streamed to client with Sovereign Visual Hierarchy
        },
      },
    },
  });

  return result.value;
}

<a name="bahasa-indonesia"></a>

Bahasa Indonesia

Integrasi Orkestrasi

Terhubung dan mengorkestrasi bersama vercel-ai-sdk-expert, senior-frontend, ui-ux-pro-max, design-system-architect, gemini-agent-booster, dan multi-agent-orchestration.

Deskripsi

UI standar bersifat statis dan di-compile sebelumnya. Di era Gemini 4 Pro, antarmuka bersifat efemeral: dihasilkan secara real-time berdasarkan niat langsung pengguna, dan dihancurkan saat tidak lagi dibutuhkan. Skill ini mengatur sistem Generative UI menggunakan React Server Components (RSC), injeksi UI berbasis Tool, dan Vercel AI SDK, memungkinkan LLM untuk langsung men-streaming komponen fungsional interaktif (grafik, formulir, dashboard) sebagai pengganti teks biasa.

Kondisi Pemicu

  • Pengguna meminta antarmuka yang beradaptasi sepenuhnya dengan konteks data (misal: "Tampilkan dashboard pengeluaran saya" -> AI menghasilkan komponen grafik kustom secara instan).
  • Mendesain antarmuka percakapan AI-native yang merender widget interaktif (formulir pemesanan, peta) alih-alih teks biasa.
  • Mengintegrasikan Vercel AI SDK streamUI atau pemetaan panggilan alat (tool-call) ke komponen React.

Arsitektur Inti (Standar 2026)

  1. Pemetaan Tool-ke-UI: AI memutuskan untuk memanggil alat (misal: getWeather). Server mencegat panggilan alat ini dan men-streaming komponen React Server (<WeatherWidget />) langsung ke klien.
  2. Zero-Client-State (Efemeral): State UI hanya ada di dalam riwayat percakapan LLM. Jika obrolan di-reset, UI ikut ter-reset.
  3. Render Optimistis: Klien langsung merender kerangka pemuatan (misal: <WeatherSkeleton />) pada milidetik pertama LLM mulai mengeluarkan token pemanggilan alat, yang kemudian diganti dengan RSC final setelah eksekusi selesai.

Kontrak Hierarki Visual pada UI Efemeral

Bahkan saat dirender secara dinamis seketika (on-the-fly), widget hasil AI WAJIB menerapkan Hierarki Visual Berdaulat:

  1. Pembatasan Kontainer: Bungkus widget streaming dalam kontainer Permukaan Level 1 (bg-white dark:bg-neutral-900 border border-neutral-200/80 dark:border-neutral-800 rounded-xl shadow-xs).
  2. Irama Tipografi: Eyebrow (text-[11px] font-semibold uppercase tracking-wider text-neutral-500) -> Judul (text-lg font-semibold tracking-tight text-neutral-900 dark:text-neutral-50) -> Teks tubuh (text-sm text-neutral-600 dark:text-neutral-300).
  3. Angka Tabular Wajib: Terapkan tabular-nums pada semua metrik angka, saldo, tarif, dan stempel waktu streaming.
  4. Dominasi Single Primary CTA: Tepat satu tombol aksi utama per kartu widget (misal: "Konfirmasi Pesanan"); aksi sekunder menggunakan gaya ghost halus.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

Expert guide for automated and manual Web Accessibility (a11y) testing — axe-core, Pa11y, Playwright a11y, screen reader testing, and WCAG 2.2 Level AA/AAA compliance / Panduan ahli pengujian aksesibilitas web.

中文简介正在准备中。当前显示原文说明。

roedyrustam/vibes-plug1052026年10月20日 更新

Expert guide for intelligent model cascading and routing — complexity-scored task routing from Flash/Haiku to Sonnet/Opus/Astra, dynamic escalation with quality gates, 40-60% token cost reduction while maintaining output quality / Panduan ahli untuk kaskade dan routing model cerdas — routing tugas berbasis skor kompleksitas dari Flash/Haiku ke Sonnet/Opus/Astra, eskalasi dinamis dengan gerbang kualitas, pengurangan biaya token 40-60% dengan kualitas output terjaga.

中文简介正在准备中。当前显示原文说明。

roedyrustam/vibes-plug1052026年10月20日 更新

Expert guide for Affective Computing, emotional AI, and real-time sentiment analysis through native multimodal tokens (voice intonation and facial micro-expressions) / Panduan ahli komputasi afektif, AI emosional, dan analisis sentimen real-time melalui token multimodal native.

中文简介正在准备中。当前显示原文说明。

roedyrustam/vibes-plug1052026年10月20日 更新

Expert guide for AI-assisted coding workflows — agentic code generation, multi-agent code swarms, self-healing CI/CD, automated PR review, spec-to-code pipelines, codebase knowledge graphs, and human-in-the-loop approval gates / Panduan ahli untuk workflow pengkodean berbasis AI — generasi kode agentic, code swarm multi-agen, CI/CD self-healing, review PR otomatis, pipeline spec-to-code, knowledge graph codebase, dan gate persetujuan human-in-the-loop.

中文简介正在准备中。当前显示原文说明。

roedyrustam/vibes-plug1052026年10月20日 更新

Expert guide for autonomous Machine-to-Machine (M2M) agent settlement, dual-rail fiat & stablecoin micro-clearing, dynamic budget envelopes, and cryptographic agent invoicing / Panduan ahli settlement otonom antar-agen (M2M), kliring mikro dual-rail fiat & stablecoin, pagu anggaran dinamis, dan penagihan agen kriptografis.

中文简介正在准备中。当前显示原文说明。

roedyrustam/vibes-plug1052026年10月20日 更新

Expert guide for long-term episodic memory integration (Mem0 v2, Letta/MemGPT, Zep v2), memory tier architecture, pgvector HNSW storage, and unified context management for autonomous AI agents / Panduan ahli untuk integrasi memori episodik jangka panjang (Mem0 v2, Letta/MemGPT, Zep v2), arsitektur tier memori, penyimpanan pgvector HNSW, dan manajemen konteks terpadu untuk agen AI otonom.

中文简介正在准备中。当前显示原文说明。

roedyrustam/vibes-plug1052026年10月20日 更新

查看 roedyrustam 的全部技能

报告此技能的问题