创意产业组 · AI Harness 行业赋能概述
1. 介绍
1.1 组的定位与范围
创意产业组覆盖六个方向:智能制造(Manufacture)、工业互联网(Industry)、媒体(Media)、创意设计(Creative)、AI 网剧(AI Drama)、AI 动画(AI Animation)。这六个方向看似分属"物质生产"与"内容生产"两个截然不同的域,但在 AI Harness 视角下它们共享同一条工程主线:把模型输出的不确定性,转化为可编排、可观测、可约束、可回归验证的交付物。
本组的划分依据不是行业分类,而是交付物的可判定性差异:
- 智能制造与工业互联网的交付物是物理世界的状态改变(一次排产、一次质检判定、一次设备检修),判定标准是良率、漏检率、停机时长等物理量,错误代价直接体现为废品与安全事故,因此 L5 评估与观测层、L6 治理与安全层是落地瓶颈。
- 媒体、创意设计、AI 网剧、AI 动画的交付物是内容资产(一篇稿、一张图、一段视频),判定标准中很大一部分是"是否还是同一个人、同一个品牌、同一个调性",这是一类跨会话、跨镜头、跨集的状态保持问题,因此 L4 记忆与状态层是质量瓶颈。
这一判断是本组全部文档的共同前提。详见 4. 六方向 × Harness 六层能力映射。
图 1-1|AI Harness 六层体系与六方向瓶颈映射
数据来源:基于本文分析绘制的示意图。
1.2 创意产业为什么需要 Harness
第一,模型能力的提升不会自动带来交付能力的提升。可口可乐 2025 年假日营销中,5 名 AI 专家在 30 天内精修超过 70,000 个视频片段(据幕后报道,非官方披露),才把生成式视频模型拉到可播出标准;而 2024 年版因人物形体扭曲受到广泛批评。这说明生成模型输出的"片段质量"与"成片质量"之间存在巨大的工程鸿沟,填补这道鸿沟的正是 Harness。
第二,合规义务已经下沉到工程细节。《人工智能生成合成内容标识办法》(国信办通字〔2025〕2 号)自 2025 年 9 月 1 日起施行,要求生成合成内容添加显式标识与隐式标识;第九条要求相关日志记录留存不少于六个月;第十条明确「任何组织和个人不得恶意删除、篡改、伪造、隐匿本办法规定的生成合成内容标识,不得为他人实施上述恶意行为提供工具或者服务」。这意味着"写元数据""打角标""留日志"不再是运营动作,而是 Harness 的 L2 工具契约与 L6 治理红线。
第三,规模已经越过了人工兜底的临界点。据中国网络视听协会《微短剧创作指引》(引自人民日报,2026),2026 年一季度全行业上线微短剧约 12.8 万部,其中 AI 微短剧约 12.2 万部,占比超 95%。在这一量级上,任何依赖人工逐条复核的质量体系都会失效,必须把一致性校验、标识校验、前置审核做成流水线上的自动关卡。
1.3 三代架构演进在本组的体现
| 代际 | 时间 | 在创意产业组的典型形态 | 失败模式 |
|---|---|---|---|
| 第一代 · 提示词工程时代 | 约 2020—2023 | 单条 Prompt 生成一张海报、一段文案;无状态、无工具 | 每次输出都要重新描述品牌与角色,风格不可复现 |
| 第二代 · 工具与编排时代 | 约 2023—2025 | 节点式工作流(ComfyUI 类)、固定流水线编排、初步资产库 | 单次任务内可控,跨会话/跨集仍漂移;效果靠人工兜底 |
| 第三代 · 运行时与评估时代 | 2025—至今 | 上下文工程 + 沙箱执行 + 轨迹评估 + 权限治理;资产库与镜头状态机成为一等公民 | 尚在成型,主要缺口是一致性指标标准化与合规自动化 |
第三代的一个关键工程原则来自 Anthropic 对 Agent Skills 的说明:让模型逐 token 生成排序结果,远比直接运行排序算法昂贵,且很多应用需要「只有代码才能提供的确定性可靠性」。对创意产业组而言,抽帧、色彩校正、元数据写入、标识校验、印刷文件导出这类操作应封装为脚本,而不是交给模型生成。这一原则贯穿本组全部 AGENTS.md 与 SKILL.md。
2. 名词解释
| 术语 | 英文/缩写 | 释义 |
|---|---|---|
| AI Harness | AI Harness | 位于基础模型与业务系统之间的工程化承载层,把模型的不确定性转化为工程可预期性 |
| 显式标识 | Explicit Label | 在文本、图片、音频、视频等可被用户明显感知的位置添加的生成合成内容提示标识 |
| 隐式标识 | Implicit Label | 嵌入文件元数据中的生成合成内容属性信息、服务提供者名称或编码、内容编号 |
| 去标识 | De-labeling | 恶意删除、篡改、伪造、隐匿生成合成内容标识的行为,《标识办法》第十条明令禁止 |
| 资产管理壳 | Asset Administration Shell, AAS | 工业 4.0 组件的数字化表达,由 Header(标识)与 Body(Submodels)构成,现行标准为 IEC 63278-1:2023 |
| 数字主线 | Digital Thread | 贯穿产品全生命周期的数据链路,ISO 23247-5:2026 定义其创建、连接、管理与维护 |
| 数字孪生 | Digital Twin | 对可观测制造元素(人员、设备、物料、工艺、设施、环境、产品、文档)的数字化表达 |
| 制造运行管理 | Manufacturing Operations Management, MOM | IEC 62264(ISA-95)五层模型中的 L3,是 IT 与 OT 的交界层 |
| 高级计划排程 | Advanced Planning and Scheduling, APS | 在多重约束下求解生产计划的排产系统,本组唐钢、思锐卓远、茗鹤案例均属此类 |
| 跨帧注意力 | Cross-frame Attention | 在视频生成中对连续帧整体建模的机制,用于抑制长序列生成中的角色漂移 |
| 智能中间帧 | Motion In-Betweening, MIB | 在关键帧之间自动生成过渡帧的技术,腾讯 VISVISE 的 4 秒 200 帧即此能力 |
| 品牌资产库 | Brand Asset Library | 承载 VI、色卡、字体、Logo、角色 IP、包装 SKU 版本的跨项目复用资产库 |
| 镜头状态机 | Shot State Machine | 以镜头为单位记录生成状态、版本、校验结果的状态机,是 AI 网剧 L4 的核心结构 |
| 人在回路 | Human-in-the-Loop, HITL | 在生成流程中设置强制人工参与节点,本组要求一致性关键节点必须前置人工确认 |
| 渐进式披露 | Progressive Disclosure | Agent Skills 的三级加载机制:元数据约 100 tokens、指令 < 5000 tokens、资源按需加载 |
3. 本组六个方向的定位对比
| 方向 | 文档 | 交付物形态 | 核心工具链 | 主导 Harness 层 | 瓶颈层 | 判定标准 |
|---|---|---|---|---|---|---|
| 智能制造 | 01-manufacture.md | 排产方案、质检判定、工艺参数 | MES / APS / SCADA / 机器视觉 / 数字孪生 | L2 + L3 | L5 + L6 | 良率、漏检率、排产达成率、OEE |
| 工业互联网 | 02-industry.md | 预警、诊断结论、工单、调度指令 | SCADA / PLC / 工单系统 / MCP 私有服务目录 / AAS | L2 + L3 | L5 + L6 | 预警提前期、误报率、MTTR、非计划停机时长 |
| 媒体 | 03-media.md | 稿件、节目、短视频、播客 | 采编系统 / 非编 / 字幕与语音合成 / 智能审核 / 分发 API | L1 + L6 | L4 + L6 | 事实准确率、审核通过率、传播效果、人工复核率 |
| 创意设计 | 04-creative.md | 平面、品牌视觉、包装、广告物料 | 文生图 / 节点式工作流 / 3D 包装打样 / 批量本地化 / 印刷文件导出 | L1 + L2 | L4 + L6 | 品牌一致性、转化率、交付周期缩短比 |
| AI 网剧 | 05-ai-drama.md | 分集成片、剧本、角色资产 | 剧本模型 / 角色库 / 分镜生成 / 视频生成 / TTS / 剪辑 | L3 + L4 | L4 | 角色一致性、风格稳定性、叙事连贯性、成本与周期 |
| AI 动画 | 06-ai-animation.md | 动画镜头、资产与序列 | 骨骼生成 / 蒙皮 / 动画生成 / 插帧 / 超分 / 调色 | L3 + L4 | L4 | ArcFace / DINO-I / CLIP-I 一致性、滑步抖动率、4K 超分质量 |
4. 六方向 × Harness 六层能力映射
说明:下表各方向的具体实现路径为基于检索资料的工程推导,表中出现的全部数值仍需回引对应方向文档的原始来源。
| 方向 | L1 上下文工程 | L2 工具与执行 | L3 编排与控制 | L4 记忆与状态 | L5 评估与观测 | L6 治理与安全 |
|---|---|---|---|---|---|---|
| 01 智能制造 | 工艺文档、设备手册向量化;ISO 23247-3 可观测制造元素属性列表可作上下文 Schema | MES / APS / 质检设备工具化;只读优先,写操作需工单语义(IEC 62264 L3 事务) | 排产—质检—运维多智能体 DAG;GB/T 39116-2020 三级"集成级"是编排的前提 | 数字主线(ISO 23247-5:2026)作为跨会话状态载体;设备孪生状态持久化(AAS / IEC 63278-1:2023) | 良率、漏检率、OEE、排产达成率;GB/T 39116-2020 四级要求"数据挖掘形成知识、模型,实现精准预测与优化" | 生产数据不出厂;核心工艺参数严禁外传;操作留痕审计 |
| 02 工业互联网 | 设备台账、图纸、维修记录;告警噪声过滤与优先级排序 | SCADA / PLC / 工单系统 MCP 化封装;反控操作必须二次确认 | 预警 → 诊断 → 备件 → 工单 闭环;试点期采用 advisory-only(仅告警)模式 | 设备健康基线(需 3~6 个月健康数据建模);运维知识引擎持续沉淀 | 预警提前期(7~15 天)、误报率(< 3%~5%)、MTTR、非计划停机时长 | 数据主权;混合云与私有化部署;IT/OT 隔离与工业协议安全边界 |
| 03 媒体 | 主流价值语料库(> 300 TB 级);事实核查检索增强;信源 grounding 是核心 | 采编 / 剪辑 / 字幕 / 审核工具集;分发平台 API | 策划 → 采集 → 生产 → 审核 → 分发 → 反馈 全链路 | 选题库、素材媒资库、已发布内容台账 | 事实准确率、审核通过率、传播效果、人工复核率 | 《标识办法》生成端显式 + 隐式标识;传播端核验与提示;不得删改伪造标识(第十条) |
| 04 创意设计 | 品牌规范(VI、色卡、字体、Logo)作为强约束上下文;参考图资产库 | 文生图 / 包装 3D 打样 / 批量本地化 / 印刷文件导出脚本 | 创意 → 草稿 → 合规校验(品牌 + 法规)→ 多区域变体 → 交付 | 品牌资产库(角色/IP/包装 SKU 版本)跨项目复用;版本与审批留痕 | 品牌一致性、转化率、生产周期缩短比例 | 训练数据版权合规;代言人与明星形象权;生成图显式标识(《标识办法》第四条第三项) |
| 05 AI 网剧 | 剧本、人物小传、角色标准三视图、分镜脚本作为锚定上下文 | 剧本模型、角色生成、分镜生成、视频生成、TTS、剪辑、BGM 各阶段工具 | 剧本 → 分镜 → 生成 → 剪辑 → 配音 五段流水线;分段生成 + 一致性校验 + 重生成 | 角色资产库(三视图 + 表情集 + 色卡 + 音色)与镜头状态机;跨集连续性 | 角色一致性(面部/服饰/道具)、风格稳定性、叙事连贯性、成本与周期 | 前置审核;显著标识;不得虚假标注主创信息;肖像权与名誉权 |
| 06 AI 动画 | 角色设定表(Character Sheet)、多角度参考图、风格参考集 | 骨骼生成 / 蒙皮 / 动画生成 / 插帧 / 超分 / 调色 各阶段模型与脚本 | 关键帧 → 中间帧 → 细节渲染 → 合成 分层流程;人为参与节点前置 | 时序记忆模块 + 跨帧注意力 + 特征传递;资产版本与镜头检查点 | 一致性指标、滑步抖动率、4K 超分质量 | 视频起始画面及播放周边显著标识(《标识办法》第四条第四项);动画形象版权;训练素材授权 |
5. 本组共性挑战与对策
| 编号 | 共性挑战 | 具体表现 | 涉及方向 | 关键层 | 对策要点 |
|---|---|---|---|---|---|
| C1 | 跨会话一致性 | 角色换脸、品牌色偏移、道具帧间不一致 | Media / Creative / AI 网剧 / AI 动画 | L4 | 建立角色与品牌资产库,以三视图 + 色卡 + 音色作为锚定上下文;把一致性做成校验关卡而非事后人工挑 |
| C2 | 标识合规自动化 | 显式标识漏打、隐式元数据在转码后丢失 | 全部含生成内容的方向 | L6 | 在导出脚本中固化元数据写入与角标渲染;导出后自动回读校验;去标识操作列为禁止事项 |
| C3 | 训练数据与素材版权 | 参考图、声音、剧本来源不清;AI 换脸明星 | Creative / AI 网剧 / AI 动画 | L6 | 建立素材授权台账;角色库仅使用授权或自有人物形象;引入前置审核 |
| C4 | 责任归属与人在回路 | AI 生成内容出现事实错误或不当表达后无人担责 | 全部方向 | L5 + L6 | 明确"AI 是执行与制作的超能力,创意方向与思想领导由人主导"的责任分工;关键节点强制人工签核 |
| C5 | 效果指标不可核验 | 大量效果数据为厂商自述口径或第三方估算 | 全部方向 | L5 | 引用时统一采用"据具体披露方(企业或机构名称)披露""第三方估算"表述;建立自有回归集与 Golden Dataset |
| C6 | 确定性操作被误交给模型 | 让模型逐帧输出调色参数、逐条写元数据 | 全部方向 | L2 | 抽帧、色彩校正、元数据写入、标识校验、印刷文件导出一律脚本化 |
| C7 | 工业数据主权与可观测性 | 产线数据不愿出网;黑箱模型不可解释 | Manufacture / Industry | L5 + L6 | 私有化或混合云部署;模型与数据分层;执行轨迹全量留痕以供追溯 |
| C8 | 标准滞后于技术 | AI 动画无专门国家/行业标准;AI 短剧规范文号未公开 | AI 网剧 / AI 动画 | L6 | 以《标识办法》为 L6 基线,行业规范作补充;无标准处如实声明"暂无" |
6. 本组文档导航
| 文件 | 内容 | 建议阅读顺序 |
|---|---|---|
| README.md | 本组概述、六方向定位对比、六层能力映射、共性挑战与对策 | 1 |
| AGENTS.md | 创意产业组上位规范:角色边界、资产与状态管理、AIGC 标识合规、版权红线、人在回路、验收标准 | 2 |
| SKILL.md | 创意产业组标准技能:一次合规的生成内容生产任务如何被拆解、执行与验收 | 3 |
| 01-manufacture.md | 智能制造:工艺、排产、质检、数字孪生 | 4 |
| 02-industry.md | 工业互联网:设备运维、能源、供应链 | 5 |
| 03-media.md | 媒体:新闻生产、内容分发、融媒体 | 6 |
| 04-creative.md | 创意设计:平面、品牌、广告、包装 | 7 |
| 05-ai-drama.md | AI 网剧 / 短剧制作流水线 | 8 |
| 06-ai-animation.md | AI 动画制作 | 9 |
AGENTS.md 与 SKILL.md 的层级关系遵循官方规格:AGENTS.md 无必需字段、纯标准 Markdown、距离被编辑文件最近者生效,可嵌套——本组 6 个方向各持有一份嵌套 AGENTS.md,组级 AGENTS.md 定义共性约束,方向级 AGENTS.md 定义专有工具链,冲突时以距离最近者为准,用户显式指令优先级最高。SKILL.md 承载任务型工作流(如"一次合规的 AIGC 物料交付"),按 name ≤ 64 字符、description ≤ 1024 字符、正文 < 5000 tokens 且 < 500 行的规格编写,并遵守渐进式披露三级加载。
7. 总结
创意产业组的六个方向在 AI Harness 体系中可以归纳为两条主线:内容侧以 L4 记忆与状态层为质量瓶颈,工业侧以 L5 评估与观测层与 L6 治理与安全层为落地瓶颈。这一差异决定了两侧的工程重心不同——内容侧要建资产库与一致性校验关卡,工业侧要建指标体系、数据主权边界与审计留痕。
两者共享的则是同一套 L6 基线:《人工智能生成合成内容标识办法》自 2025 年 9 月 1 日施行,显式与隐式标识、去标识日志留存不少于六个月、禁止删改伪造标识,已经把合规义务下沉到元数据字段与导出脚本这一级。任何创意产业方向的 Harness 设计,如果不能回答"这条内容的标识写在哪里、由谁写、写完怎么验",就还不具备交付条件。
需要强调的是,本组引用的大量效果数据来自厂商案例页、券商研报汇编、第三方估算与自媒体复盘,可信度层级不一。本文档工程统一采用"据具体披露方(企业或机构名称)披露""第三方估算""非官方披露"等表述加以标注,不将其提升为断言,并在各方向文档末尾设置信息缺口声明。
信息缺口声明
| 缺口项 | 处理方式 |
|---|---|
| 《网络安全技术 人工智能生成合成内容标识方法》的 GB 编号 | 已确认:GB 45438—2025(强制性国标,2025-02-28 发布、2025-09-01 实施,与《标识办法》同步;来源:国家标准全文公开系统、TC260 官方文本),不臆造编号 |
| 《标识办法》配套的 6 项网络安全标准实践指南全称 | 仅确认其中《人工智能生成合成内容标识方法 文件元数据隐式标识 文本文件》,其余 5 项名称未检索到 → 不列全称 |
| 《网络视听 AI 短剧创作生产与传播规范》文号与条款全文 | 仅检索到二手页面称 2025 年 4 月由中国网络视听节目服务协会发布 → 只写时间与机构,标注 |
| GB/T 39116-2020 能力子域口径 | 两种口径并存 → 本组统一采用 PTRM 四要素 / 8 能力域 / 20 能力子域,并在文中注明另一口径 |
| AI 动画的国家/行业标准 | 未检索到专门的 AI 动画国家标准或行业标准 → 如实声明"暂无" |
8. 参考资料
- 人工智能生成合成内容标识办法 — 国家互联网信息办公室、工业和信息化部、公安部、国家广播电视总局,2025。https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
- 多措并举推进标识体系建设,助力新时代人工智能健康发展 — 国家互联网应急中心,2025。https://www.cac.gov.cn/2025-09/06/c_1758880709361356.htm
- 《人工智能生成合成内容标识办法》助力辨别虚假信息 — 中国政府网,2025。https://www.gov.cn/zhengce/202503/content_7014404.htm
- 9月1日起 AI生成合成内容必须添加标识 — 央视网,2025。https://big5.cctv.com/gate/big5/news.cctv.cn/2025/09/01/ARTI3ZlXK7MyM39Pm3PuZ5Hm250901.shtml
- 拥抱智能浪潮 泳向变革深处 — 新华社,2025。https://www.news.cn/20251114/1c01598d836c449bbfd54267b6ecea6d/c.html
- 主流媒体所办新媒体发展研究报告(2024-2025)— 人民网,2025。https://sc.people.com.cn/BIG5/n2/2025/1030/c345167-41396739.html
- AI 要给微短剧"洗牌"?— 人民日报,2026。https://kpzg.people.com.cn/n1/2026/0511/c404214-40717117.html
- AGENTS.md 官方站 — Agentic AI Foundation(Linux Foundation),2025—2026。https://agents.md/
- Lesson 15: AGENTS.md - giving agents project context — Addy Osmani,2026。https://addyosmani.com/agents/15-agents-md/
- AGENTS.md Boilerplate(写作原则)— Sentry 开发者文档。https://develop.sentry.dev/sdk/getting-started/templates/agents-md
- Agent Skills Specification — agentskills.io,2025—2026。https://agentskills.io/specification
- Equipping agents for the real world with Agent Skills — Anthropic,2025(2025-12-18 更新为开放标准)。https://claude.com/blog/equipping-agents-for-the-real-world-with-agent-skills
- Building custom Skills for Claude — Anthropic Claude Cookbook,2025。https://platform.claude.com/cookbook/skills-notebooks-03-skills-custom-development
- 智能制造能力成熟度评估介绍(GB/T 39116-2020、GB/T 39117-2020)— 中国电子技术标准化研究院。https://www.cc.cesi.cn/service/show-2478.aspx
- ISO 23247-1:2021 Digital twin framework for manufacturing — ISO。https://www.iso.org/standard/75066.html
- Coca-Cola Refreshes Givers of the Season — The Coca-Cola Company,2025。https://coca-colacompany.com/media-center/coca-cola-refreshes-givers-of-the-season-embraces-ai-powered-storytelling-in-global-holiday-campaign
Creative Industry Group · AI Harness Industry Enablement Overview
1. Introduction
1.1 Positioning and Scope of the Group
The Creative Industry Group covers six directions: Smart Manufacturing, Industrial Internet, Media, Creative Design, AI Drama, and AI Animation. These six directions may seem to belong to two entirely different domains — "material production" and "content production" — but from the AI Harness perspective they share one common engineering thread: converting the uncertainty of model output into deliverables that are orchestratable, observable, constrainable, and verifiable through regression.
This group's division is based not on industry classification, but on the difference in adjudicability of the deliverables:
- For Smart Manufacturing and Industrial Internet, the deliverables are state changes in the physical world (a production scheduling run, a quality-inspection decision, a piece of equipment maintenance), and the judging criteria are physical quantities such as yield, miss rate, and downtime. The cost of errors is directly reflected as waste and safety incidents, so the L5 evaluation and observation layer and the L6 governance and safety layer are the deployment bottleneck.
- For Media, Creative Design, AI Drama, and AI Animation, the deliverables are content assets (an article, an image, a video), and a large part of the judging criteria concerns "whether it is still the same person, the same brand, the same tone." This is a class of cross-session, cross-shot, cross-episode state-keeping problems, so the L4 memory and state layer is the quality bottleneck.
This judgment is the common premise of all documents in this group. See 4. Six Directions × Harness Six-Layer Capability Mapping.
图 1-1|AI Harness 六层体系与六方向瓶颈映射
数据来源:基于本文分析绘制的示意图。
1.2 Why the Creative Industry Needs Harness
First, improvements in model capability do not automatically translate into improvements in delivery capability. In Coca-Cola's 2025 holiday campaign, 5 AI experts spent 30 days fine-tuning more than 70,000 video clips (per behind-the-scenes reports, not officially disclosed) before the generative video model reached a broadcastable standard; the 2024 version, by contrast, drew widespread criticism for distorted figures. This shows a huge engineering gap between the "clip quality" of generative model output and the "final-film quality," and it is precisely this gap that Harness fills.
Second, compliance obligations have already been pushed down to engineering details. The "Measures for the Labeling of AI-Generated and Synthesized Content" (National Cyberspace Administration Protocol No. [2025] 2) took effect on September 1, 2025, requiring generated and synthesized content to carry explicit and implicit labels; Article 9 requires relevant logs to be retained for no less than six months; Article 10 specifies that "no organization or individual may maliciously delete, alter, forge, or conceal the labels for generated and synthesized content specified in these Measures, nor provide tools or services for others to carry out the above malicious acts." This means that "writing metadata," "adding corner marks," and "keeping logs" are no longer operational tasks, but rather the L2 tool contract and the L6 governance red line for Harness.
Third, scale has already crossed the inflection point where manual fallback is feasible. According to the China Netcasting Services Association's "Micro-Drama Creation Guide" (cited from People's Daily, 2026), in Q1 2026 the industry launched about 128,000 micro-dramas, of which about 122,000 were AI micro-dramas, accounting for over 95%. At this scale, any quality system relying on manual item-by-item review will fail; consistency checks, labeling checks, and pre-review must be turned into automated gates on the production line.
1.3 Manifestation of the Three-Generation Architecture Evolution in this Group
| Generation | Time | Typical Form in the Creative Industry Group | Failure Mode |
|---|---|---|---|
| 1st Gen · Prompt Engineering Era | c. 2020—2023 | A single Prompt generates a poster or a piece of copy; stateless, tool-free | Brand and character must be re-described on every output; style is not reproducible |
| 2nd Gen · Tooling and Orchestration Era | c. 2023—2025 | Node-based workflows (ComfyUI-like), fixed-pipeline orchestration, initial asset libraries | Controllable within a single task, but still drifts across sessions/episodes; results depend on manual fallback |
| 3rd Gen · Runtime and Evaluation Era | 2025—present | Context engineering + sandboxed execution + trajectory evaluation + permission governance; asset libraries and shot state machines become first-class citizens | Still taking shape; the main gaps are standardization of consistency metrics and compliance automation |
A key engineering principle of the third generation comes from Anthropic's explanation of Agent Skills: having a model generate sorted results token by token is far more expensive than directly running a sorting algorithm, and many applications need "the deterministic reliability that only code can provide." For the Creative Industry Group, operations such as frame extraction, color correction, metadata writing, label verification, and print-file export should be wrapped as scripts rather than handed to the model. This principle runs through all the AGENTS.md and SKILL.md files in this group.
2. Glossary
| Term | English / Abbreviation | Definition |
|---|---|---|
| AI Harness | AI Harness | The engineered carrier layer located between the foundation model and the business system, converting model uncertainty into engineering predictability |
| Explicit Label | Explicit Label | A prompt label for generated and synthesized content added at positions in text, images, audio, video, etc. that users can perceive clearly |
| Implicit Label | Implicit Label | Attribute information of generated and synthesized content, the service provider's name or code, and the content number embedded in file metadata |
| De-labeling | De-labeling | The act of maliciously deleting, altering, forging, or concealing the labels of generated and synthesized content, expressly prohibited by Article 10 of the Labeling Measures |
| Asset Administration Shell | Asset Administration Shell, AAS | The digital expression of an Industry 4.0 component, composed of a Header (identification) and a Body (Submodels); the current standard is IEC 63278-1:2023 |
| Digital Thread | Digital Thread | The data chain running through the full product lifecycle; ISO 23247-5:2026 defines its creation, connection, management, and maintenance |
| Digital Twin | Digital Twin | The digital expression of observable manufacturing elements (personnel, equipment, materials, processes, facilities, environment, products, documents) |
| Manufacturing Operations Management | Manufacturing Operations Management, MOM | L3 in the five-layer model of IEC 62264 (ISA-95); the boundary layer between IT and OT |
| Advanced Planning and Scheduling | Advanced Planning and Scheduling, APS | A production-scheduling system that solves production plans under multiple constraints; the Tangsteel, SIRUI Zhuoyuan, and Minghe cases in this group all fall into this category |
| Cross-frame Attention | Cross-frame Attention | A mechanism in video generation that models consecutive frames as a whole, used to suppress character drift in long-sequence generation |
| Motion In-Betweening | Motion In-Betweening, MIB | A technique for automatically generating transition frames between keyframes; Tencent VISVISE's 200 frames in 4 seconds is this capability |
| Brand Asset Library | Brand Asset Library | A cross-project reusable asset library carrying VI, color cards, fonts, logos, character IP, and packaging SKU versions |
| Shot State Machine | Shot State Machine | A state machine that records generation state, versions, and verification results on a per-shot basis; the core structure of the L4 layer in AI Drama |
| Human-in-the-Loop | Human-in-the-Loop, HITL | Mandatory human-participation nodes set up in the generation process; this group requires manual confirmation up front at consistency-critical nodes |
| Progressive Disclosure | Progressive Disclosure | The three-level loading mechanism of Agent Skills: metadata ~100 tokens, instructions < 5000 tokens, resources loaded on demand |
3. Positioning Comparison of the Group's Six Directions
| Direction | Document | Deliverable Form | Core Toolchain | Dominant Harness Layer | Bottleneck Layer | Judging Criteria |
|---|---|---|---|---|---|---|
| Smart Manufacturing | 01-manufacture.md | Production scheduling plans, quality-inspection decisions, process parameters | MES / APS / SCADA / Machine Vision / Digital Twin | L2 + L3 | L5 + L6 | Yield, miss rate, scheduling attainment rate, OEE |
| Industrial Internet | 02-industry.md | Alerts, diagnostic conclusions, work orders, dispatch instructions | SCADA / PLC / Work-order system / MCP private service catalog / AAS | L2 + L3 | L5 + L6 | Alert lead time, false-alarm rate, MTTR, unplanned downtime |
| Media | 03-media.md | Manuscripts, programs, short videos, podcasts | Editing/acquisition system / Non-linear editing / Subtitles and speech synthesis / Intelligent review / Distribution API | L1 + L6 | L4 + L6 | Factual accuracy, review pass rate, distribution effect, manual re-review rate |
| Creative Design | 04-creative.md | Print design, brand visuals, packaging, advertising materials | Text-to-image / Node-based workflows / 3D packaging mockup / Batch localization / Print-file export | L1 + L2 | L4 + L6 | Brand consistency, conversion rate, delivery-cycle shortening ratio |
| AI Drama | 05-ai-drama.md | Finished episodes, scripts, character assets | Script model / Character library / Storyboard generation / Video generation / TTS / Editing | L3 + L4 | L4 | Character consistency, style stability, narrative coherence, cost and cycle |
| AI Animation | 06-ai-animation.md | Animation shots, assets and sequences | Rig generation / Skinning / Animation generation / Interpolation / Super-resolution / Color grading | L3 + L4 | L4 | ArcFace / DINO-I / CLIP-I consistency, sliding-jitter rate, 4K super-resolution quality |
4. Six Directions × Harness Six-Layer Capability Mapping
Note: The specific implementation path of each direction in the table below is an engineering derivation based on retrieved materials; all values appearing in the table still need to be traced back to the original sources of the corresponding direction documents.
| Direction | L1 Context Engineering | L2 Tools and Execution | L3 Orchestration and Control | L4 Memory and State | L5 Evaluation and Observation | L6 Governance and Safety |
|---|---|---|---|---|---|---|
| 01 Smart Manufacturing | Vectorization of process documents and equipment manuals; the ISO 23247-3 list of observable manufacturing element attributes can serve as the context Schema | Tooling of MES / APS / inspection equipment; read-only first, write operations require work-order semantics (IEC 62264 L3 transactions) | Scheduling—inspection—operations multi-agent DAG; the Level-3 "integration level" of GB/T 39116-2020 is the prerequisite for orchestration | Digital Thread (ISO 23247-5:2026) as the cross-session state carrier; persistence of equipment twin state (AAS / IEC 63278-1:2023) | Yield, miss rate, OEE, scheduling attainment rate; the Level-4 requirement of GB/T 39116-2020 calls for "data mining to form knowledge and models, enabling accurate prediction and optimization" | Production data does not leave the plant; core process parameters are strictly forbidden from being transmitted externally; operations are logged and audited |
| 02 Industrial Internet | Equipment ledgers, drawings, and maintenance records; alert-noise filtering and priority ranking | MCP-style wrapping of SCADA / PLC / work-order systems; reverse-control operations require second confirmation | Alert → Diagnosis → Spare parts → Work order closed loop; the pilot phase adopts an advisory-only (alert-only) mode | Equipment health baseline (requires 3–6 months of health-data modeling); continuous accumulation of an operations knowledge engine | Alert lead time (7–15 days), false-alarm rate (< 3%–5%), MTTR, unplanned downtime | Data sovereignty; hybrid-cloud and on-premises deployment; IT/OT isolation and industrial-protocol security boundaries |
| 03 Media | Mainstream-value corpus (> 300 TB class); fact-checking retrieval augmentation; source grounding is the core | Tool suite for acquisition/editing / video editing / subtitles / review; distribution platform APIs | Planning → Gathering → Production → Review → Distribution → Feedback end-to-end chain | Topic library, material/media-asset library, published-content ledger | Factual accuracy, review pass rate, distribution effect, manual re-review rate | Explicit + implicit labels on the generation side per the Labeling Measures; verification and prompting on the distribution side; no deleting/altering/forging of labels (Article 10) |
| 04 Creative Design | Brand specifications (VI, color cards, fonts, logos) as strongly constraining context; reference-image asset library | Text-to-image / packaging 3D mockup / batch localization / print-file export scripts | Creative → Draft → Compliance check (brand + regulations) → Multi-region variants → Delivery | Brand asset library (character/IP/packaging SKU versions) reused across projects; version and approval trail | Brand consistency, conversion rate, production-cycle shortening ratio | Training-data copyright compliance; endorsement and celebrity image rights; explicit labels on generated images (Article 4, Item 3 of the Labeling Measures) |
| 05 AI Drama | Scripts, character profiles, character standard three-view sheets, and storyboard scripts as anchoring context | Per-stage tools for script model, character generation, storyboard generation, video generation, TTS, editing, BGM | Script → Storyboard → Generation → Editing → Dubbing five-stage pipeline; segment-based generation + consistency check + regeneration | Character asset library (three-view sheets + expression sets + color cards + voice timbre) and shot state machine; cross-episode continuity | Character consistency (face/wardrobe/props), style stability, narrative coherence, cost and cycle | Pre-review; prominent labels; no false attribution of creator credits; portrait rights and reputation rights |
| 06 AI Animation | Character sheets, multi-angle reference images, style reference sets | Per-stage models and scripts for rig generation / skinning / animation generation / interpolation / super-resolution / color grading | Keyframe → In-between → Detail rendering → Compositing layered workflow; human-participation nodes moved up front | Temporal memory module + cross-frame attention + feature propagation; asset versions and shot checkpoints | Consistency metrics, sliding-jitter rate, 4K super-resolution quality | Prominent labels on the video start frame and around playback (Article 4, Item 4 of the Labeling Measures); authorship rights of animated characters; training-material licensing |
5. Common Challenges and Countermeasures
| No. | Common Challenge | Specific Manifestation | Directions Involved | Key Layer | Countermeasure Key Points |
|---|---|---|---|---|---|
| C1 | Cross-session consistency | Character face-swapping, brand-color drift, inconsistent props across frames | Media / Creative / AI Drama / AI Animation | L4 | Build character and brand asset libraries, using three-view sheets + color cards + voice timbre as anchoring context; make consistency a verification gate rather than a post-hoc manual pick |
| C2 | Label-compliance automation | Missed explicit labels; implicit metadata lost after transcoding | All directions containing generated content | L6 | Embed metadata writing and corner-mark rendering in export scripts; automatically re-read and verify after export; classify de-labeling as a prohibited act |
| C3 | Training-data and material copyright | Unclear provenance of reference images, voices, and scripts; AI face-swapping of celebrities | Creative / AI Drama / AI Animation | L6 | Build a material-licensing ledger; character libraries use only licensed or owned likenesses; introduce pre-review |
| C4 | Accountability and human-in-the-loop | No one is held responsible when AI-generated content contains factual errors or improper expression | All directions | L5 + L6 | Clarify the responsibility split that "AI is the superpower for execution and production, while creative direction and thought leadership are led by humans"; enforce human sign-off at critical nodes |
| C5 | Effect metrics not verifiable | Much effect data comes from vendor self-reported statements or third-party estimates | All directions | L5 | When citing, uniformly adopt phrasings such as "as disclosed by the specific disclosing party (company or institution name)" and "third-party estimate"; build an in-house regression set and a Golden Dataset |
| C6 | Deterministic operations mistakenly handed to the model | Having the model output color-grading parameters frame by frame and write metadata item by item | All directions | L2 | Frame extraction, color correction, metadata writing, label verification, and print-file export are always scripted |
| C7 | Industrial data sovereignty and observability | Production-line data is kept off the network; black-box models are not explainable | Manufacture / Industry | L5 + L6 | On-premises or hybrid-cloud deployment; layer the model and the data; keep full execution-trail records for traceability |
| C8 | Standards lag behind technology | No dedicated national/industry standard for AI animation; the document number of the AI micro-drama specification is not public | AI Drama / AI Animation | L6 | Use the Labeling Measures as the L6 baseline, with industry specifications as supplements; where no standard exists, honestly state "none" |
6. Document Navigation for this Group
| File | Content | Recommended Reading Order |
|---|---|---|
| README.md | Overview of this group, positioning comparison of the six directions, six-layer capability mapping, common challenges and countermeasures | 1 |
| AGENTS.md | Top-level specification of the Creative Industry Group: role boundaries, asset and state management, AIGC label compliance, copyright red lines, human-in-the-loop, acceptance criteria | 2 |
| SKILL.md | Standard skills of the Creative Industry Group: how a compliant generated-content production task is decomposed, executed, and accepted | 3 |
| 01-manufacture.md | Smart Manufacturing: processes, scheduling, inspection, digital twin | 4 |
| 02-industry.md | Industrial Internet: equipment operations, energy, supply chain | 5 |
| 03-media.md | Media: news production, content distribution, converged media | 6 |
| 04-creative.md | Creative Design: print, brand, advertising, packaging | 7 |
| 05-ai-drama.md | AI Drama / short-drama production pipeline | 8 |
| 06-ai-animation.md | AI animation production | 9 |
The hierarchical relationship between AGENTS.md and SKILL.md follows the official specification: AGENTS.md has no required fields, is plain standard Markdown, the closest one to the file being edited takes effect, and it can be nested — each of this group's six directions holds its own nested AGENTS.md; group-level AGENTS.md defines common constraints while direction-level AGENTS.md defines dedicated toolchains; in the event of conflict, the closest one prevails, and explicit user instructions have the highest priority. SKILL.md carries task-oriented workflows (e.g. "a compliant AIGC material delivery"), written to the specification of a name ≤ 64 characters, a description ≤ 1024 characters, and a body < 5000 tokens and < 500 lines, while following the three-level load of progressive disclosure.
7. Summary
The six directions of the Creative Industry Group can be summarized into two main threads within the AI Harness system: on the content side, the L4 memory and state layer is the quality bottleneck; on the industrial side, the L5 evaluation and observation layer and the L6 governance and safety layer are the deployment bottleneck. This difference determines that the two sides have different engineering priorities — the content side must build asset libraries and consistency-check gates, while the industrial side must build an indicator system, data-sovereignty boundaries, and audit trails.
What the two sides share is the same set of L6 baselines: the "Measures for the Labeling of AI-Generated and Synthesized Content" took effect on September 1, 2025; explicit and implicit labels, de-labeling log retention of no less than six months, and the prohibition on deleting/altering/forging labels have already pushed compliance obligations down to the level of metadata fields and export scripts. Any Harness design in a creative-industry direction that cannot answer "where is this content's label written, who writes it, and how is it verified afterward" is not yet ready for delivery.
It should be stressed that much of the effect data cited in this group comes from vendor case pages, broker-research compilations, third-party estimates, and self-media retrospectives, with varying levels of credibility. This document uniformly uses phrasings such as "as disclosed by the specific disclosing party (company or institution name)," "third-party estimate," and "non-official disclosure" to annotate such data rather than elevating it to assertions, and sets up an information-gap declaration at the end of each direction document.
Information Gap Declaration
| Gap Item | Handling Approach |
|---|---|
| The GB number of "Network Security Technology — Labeling Methods for AI-Generated and Synthesized Content" | Confirmed: GB 45438—2025 (a mandatory national standard, published on 2025-02-28 and effective on 2025-09-01, in sync with the Labeling Measures; source: the national-standard full-text disclosure system and the official TC260 text); no invented number |
| Full names of the 6 network-security standard practice guides supporting the Labeling Measures | Only "Labeling Methods for AI-Generated and Synthesized Content — Implicit Labeling of File Metadata — Text Files" is confirmed; the names of the other 5 have not been found → do not list full names |
| Document number and full text of the "Code of Practice for the Creation, Production, and Distribution of AI Short Dramas over Network Audiovisual Services" | Only a second-hand page stating it was released in April 2025 by the China Netcasting Services Association was found → write only the time and the organization, and mark it [To be verified] |
| Caliber of the capability sub-domains of GB/T 39116-2020 | Two calibers coexist → this group uniformly adopts the PTRM four elements / 8 capability domains / 20 capability sub-domains, and notes the other caliber in the text |
| National/industry standards for AI animation | No dedicated national or industry standard for AI animation was found → honestly state "none" |
8. References
- Measures for the Labeling of AI-Generated and Synthesized Content — Cyberspace Administration of China, Ministry of Industry and Information Technology, Ministry of Public Security, National Radio and Television Administration, 2025. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
- Promoting the Construction of the Labeling System through Multiple Measures to Support the Healthy Development of AI in the New Era — National Internet Emergency Center, 2025. https://www.cac.gov.cn/2025-09/06/c_1758880709361356.htm
- "Measures for the Labeling of AI-Generated and Synthesized Content" Helps Identify False Information — Chinese Government Website, 2025. https://www.gov.cn/zhengce/202503/content_7014404.htm
- From September 1, AI-Generated and Synthesized Content Must Carry Labels — CCTV, 2025. https://big5.cctv.com/gate/big5/news.cctv.cn/2025/09/01/ARTI3ZlXK7MyM39Pm3PuZ5Hm250901.shtml
- Embracing the Intelligent Wave and Swimming Toward the Depths of Change — Xinhua News Agency, 2025. https://www.news.cn/20251114/1c01598d836c449bbfd54267b6ecea6d/c.html
- Research Report on New Media Run by Mainstream Media (2024–2025) — People's Daily Online, 2025. https://sc.people.com.cn/BIG5/n2/2025/1030/c345167-41396739.html
- Will AI "Shuffle the Deck" for Micro-Dramas? — People's Daily, 2026. https://kpzg.people.com.cn/n1/2026/0511/c404214-40717117.html
- AGENTS.md official site — Agentic AI Foundation (Linux Foundation), 2025—2026. https://agents.md/
- Lesson 15: AGENTS.md - giving agents project context — Addy Osmani, 2026. https://addyosmani.com/agents/15-agents-md/
- AGENTS.md Boilerplate (writing principles) — Sentry developer documentation. https://develop.sentry.dev/sdk/getting-started/templates/agents-md
- Agent Skills Specification — agentskills.io, 2025—2026. https://agentskills.io/specification
- Equipping agents for the real world with Agent Skills — Anthropic, 2025 (updated to an open standard on 2025-12-18). https://claude.com/blog/equipping-agents-for-the-real-world-with-agent-skills
- Building custom Skills for Claude — Anthropic Claude Cookbook, 2025. https://platform.claude.com/cookbook/skills-notebooks-03-skills-custom-development
- Introduction to the Assessment of Smart Manufacturing Capability Maturity (GB/T 39116-2020, GB/T 39117-2020) — China Electronics Standardization Institute. https://www.cc.cesi.cn/service/show-2478.aspx
- ISO 23247-1:2021 Digital twin framework for manufacturing — ISO. https://www.iso.org/standard/75066.html
- Coca-Cola Refreshes Givers of the Season — The Coca-Cola Company, 2025. https://coca-colacompany.com/media-center/coca-cola-refreshes-givers-of-the-season-embraces-ai-powered-storytelling-in-global-holiday-campaign