AdDesign-Agent is a Codex-powered automated commercial graphic advertising designer. It turns rough customer requests into structured design briefs, copy hierarchy, layout plans, image-generation prompts, QA reports, print specs, revision instructions, and final delivery package structures.
The project is designed for practical advertising work: signboard lightboxes, posters, price lists, culture walls, safety boards, signature walls, stickers, menus, logo avatars, promotional graphics, and campaign images.
Open this repository in Codex and describe the customer request in plain language. The project-level AGENTS.md tells Codex to use the local skills under skills/ in the correct order:
- Understand the request.
- Parse a structured brief.
- Plan copy and layout.
- Build image prompts.
- QA the design.
- Produce print and delivery guidance.
You can enter incomplete, conversational requests:
设计一张清吧门头灯箱，尺寸横版3.5x1.6米，内容：蓝熏小酒窝，广告语：有酒有局，便有了江湖！风格：美式复古工业风，要求图案简单，不复杂，背景不要太暗，生成4K高清图片，平面二维图。
The Agent should output:
- Structured brief
- Layout plan
- Chinese image-generation prompt
- English image-generation prompt
- Negative prompt
- QA checklist
- Print delivery recommendation
brand-brief-parser extracts design intent, size, text, constraints, and missing information.
ad-copywriter preserves customer text and organizes title, slogan, body copy, price, and warning layers.
industry-style-library selects visual language by industry.
layout-planner turns the brief into composition and element placement.
asset-ingestion defines how logos, QR codes, product images, people, and reference images should be handled.
chinese-typography checks Chinese readability, font direction, scale, and punctuation.
image-prompt-builder creates Chinese and English prompts plus negative prompts.
design-qa checks correctness before generation and delivery.
revision-controller translates feedback into precise edit instructions.
print-production gives size, pixel, bleed, safe-margin, and export guidance.
delivery-packager creates the expected final handoff structure.
When the user says things like "太黄了", "太复杂了", "文字太小", "二维码醒目一点", or "不要生硬地放进去", the Agent first classifies the feedback type, then updates only the affected prompt, layout, style, asset, or print rules while preserving approved content.
Use:
python scripts/delivery_packager.py --root . --output deliveryThe script creates the delivery folders, checks expected files, and writes delivery/delivery_summary.md.
For the bar signboard example, Codex should produce a brief with:
project_type: 门头灯箱industry: 清吧size_real_world: 3.5 x 1.6 maspect_ratio: 35:16main_text: 蓝熏小酒窝slogan: 有酒有局，便有了江湖！hard_constraints: 图案简单、不复杂、不凌乱、背景不要太暗、平面二维图、4K高清
Then it should produce a horizontal signboard layout, flat 2D prompt, negative prompt focused on Chinese text fidelity, QA checklist, and practical lightbox print notes.
Add new skills for specific industries, new layout reference files, deterministic production scripts, or example briefs for recurring customers. Keep skills concise and store detailed knowledge in references/ files so Codex can load only what it needs.