Vibe Coding: Building an AI-Powered Creative Platform for Consumer-Facing Design

Introduction

C-end visual demands exhibit strong productization and customization characteristics: business requests are scattered and high-volume, while traditional template-based AI design struggles to balance efficiency and quality. Unstable model calls, non-standard outputs, and high manual intervention costs are the primary bottlenecks in current C-end visual production.

To address these issues, the team adopted Vibe Coding as the primary development approach and built an AI image-generation Skill workflow platform for C-end business. The platform automates the full chain from requirement parsing and model scheduling to image-generation iteration and material output.

Architecture Design

The platform skeleton is a co-creatable, closed-loop workflow system centered on three core links:

Requirement Parsing & Distribution: Analyzes business inputs (screenshots, PRDs, descriptive text) and routes them to the most suitable model and workflow.

Efficient Image Generation & Auto-Iteration: Leverages the group's AI image-generation infrastructure, combined with quality auditing and case-library mechanisms, to optimize generation paths.

Low-Cost Usage: Standardizes output files and embeds the tool into existing workflows via seamless interactions such as DingTalk bots.

Core Capabilities

1. Intelligent Prompt Polishing

Drawing on Vibe design thinking, the system automatically parses brief user input and generates structured prompts based on historical generation experience. This front-loads the most time-consuming "requirement translation" step, turning it into a standardized, predictable process that significantly improves output stability.

2. Vibe/Master Dual Modes

Differentiated design for two core user groups — professional designers and business users — enables seamless switching within the same platform.

Vibe Mode: Supports rapid generation and creative exploration; non-designers can produce complete atmospheric visuals from a single sentence or screenshot.

Master Mode: Provides fine-grained control and deep tuning for professionals, with multi-dimensional preset parameters before generation and ample post-generation editing space, balancing lightweight speed with high-precision creation.

3. Advanced Material Style Library

Based on high-frequency project scenarios, the platform presets 10+ material libraries and 10+ style libraries. Users select target materials and styles, adjust texture intensity, and generate quickly without repeatedly uploading references or rewriting prompts, boosting both efficiency and visual consistency.

4. Fine-Grained Secondary Editing

For post-generation detail adjustments, the platform offers quick commands and natural-language-driven local modifications: smart HD, proportional expansion, background removal, smart color adjustment, and PSD downloads for selected design points.

5. Full-Chain Efficient Expansion

Beyond single design points, the platform supports standardized presets for multi-end adaptation. Starting from one design scheme, it rapidly produces outputs for multiple endpoints with different specifications, overturning the traditional sequential manual adaptation process.

Practical Applications

1. Free Creative Image Generation

For non-standard needs — creative divergence, main-visual exploration, illustration assets, effect images — the platform's rich material and style presets enable efficient exploration. Before formal generation, the workflow polishes prompts and supports multi-style output in one run, ideal for early-stage visual ideation.

Example: Based on an existing boot-screen specification, a brief prompt quickly generated four seasonal design variants with near-production preview fidelity.

2. Business Custom Workflow Image Generation

Targeting high-volume operational needs (repeated reskins, UI-display cards), the team built a dedicated end-to-end pipeline: creative input → AI generation → rendering composition → slicing output → content review — all one-click. Outputs are sliced into customized assets (precise background images, logo slices, color values) and packaged in a folder for direct business configuration.

Prompt Engineering Layer: Each card type has an independent prompt template system supporting 3-style parallel generation. The "Flash Purchase Knight" IP is controlled via detailed character descriptions + reference images (helmet, goggles, bamboo-copter, body proportions). Negative constraints ("no text/watermark/logo/border") are auto-appended per scenario to prevent unwanted elements.

Scene Style Engine (Popup Example): A three-dimensional parameter matrix — 10 audiences × 11 scenes × 3 visual levels — yields 330+ differentiated prompt combinations; operators simply pick three tags for auto-matching.

Rendering Composition Layer: AI raw image → brand-color gradient → logo/copy/tag overlay → rounded crop. Text uses 3× supersampling + shear rotation + LANCZOS scaling to eliminate transformation artifacts.

Results & Outlook

Measured project statistics show single-image demand processing time compressed from 2–3 hours to under 3–5 minutes. The platform embeds a quality-audit, case-library, and preference-analysis loop; high-quality outputs continuously feed back into the prompt generator, aligning the tool with evolving business aesthetic preferences.

The core value lies not in a single technical breakthrough but in establishing a sustainable "Business–Design–Technology" tripartite collaboration mechanism :

Business contributes product demands and participates in design-material co-creation at low threshold.

UED defines rules and aesthetic standards; designers encapsulate Skill modules and precipitate a knowledge base.

Platform shoulders efficient execution and learning of design volume.

This effort reconstructs C-end visual production from a labor-intensive process into a business-driven intelligent pipeline. Future iterations will extend fine-grained creative functions and broaden business-scenario coverage.