Four task levels decide how much process a request deserves四个任务等级决定一个请求值得多少流程

The router reads intent and assigns a level. L1 simple image: no copy, DNA optional. L2 main image or scene image: optional selling-point tags, DNA recommended. L3 detail-page poster: copy required for every frame, DNA required. L4 creative poster: marketing copy required. Copy skills are only invoked when the level demands them, which is what made the system fast for the common case and rigorous for the expensive one.路由器识别意图并分配等级。L1 简单图片：无文案，DNA 可选。L2 主图或场景图：可选卖点标签，推荐 DNA。L3 商详海报：每一帧必须有文案，必须有 DNA。L4 创意海报：必须有营销文案。文案 skill 只在等级需要时才被调用，这让系统在常见情况下快、在昂贵情况下严。

DNA generators: extract when there are samples, generate when there aren'tDNA 生成器：有样本就提取，没样本就生成

Two generators, one for copy DNA and one for design DNA, produce the same standardized document either way. Given competitor links, uploaded images, past copy or a brand VI file, they extract patterns. Given only brand information and product traits, they generate a starting DNA. Every downstream skill reads the same format, so the agent's "taste" for a brand is a file that can be reviewed and edited by a human.两个生成器，一个负责文案 DNA，一个负责设计 DNA，无论哪种模式都产出同一种标准化文档。给它竞品链接、上传图片、历史文案或品牌 VI 文件，它提取模式；只给品牌信息和产品特性，它生成一份起始 DNA。所有下游 skill 读同一种格式，于是 Agent 对一个品牌的「品味」是一个人可以审阅和修改的文件。

Identity library: a product database and a prompt database身份库：一个商品库，一个提示词库

The product DB stores each product's shape, visual assets and feature-advantage-benefit copy. The prompt DB stores templates by task type × category. Together they are the agent's memory, so the tenth image of a product is cheaper and more consistent than the first.商品库存每个商品的外形、视觉资产和 FAB 文案；提示词库按「任务类型 × 品类」存模板。两者合起来是 Agent 的记忆，让一个商品的第十张图比第一张更便宜、更一致。

Identity-preservation rules: what the model may never invent身份保持规则：模型绝不允许编造的东西

• →Product shape must be preserved: prefer image-to-image or cutout compositing over generation.商品外形必须保持：优先图生图或抠图合成，而不是生成。
• →Fixed models (people) must be preserved when a brand has them.品牌有固定模特时，模特外形必须保持。
• →Logos are overlaid in post; the model never draws them.Logo 后期叠加，模型不画。
• →Text is typeset in post; the model never renders copy.文字后期排版，模型不渲染文案。