AI Harness for AI Drama

1. Introduction

1.1 Background

AI dramas (AI micro-dramas) are the direction within the Creative Industries cluster that is scaling the fastest while being the most lagging in specification-building. According to the China Netcasting Services Association's Guidelines for Micro-drama Creation (as cited in People's Daily, 2026), in Q1 2026 the industry went live with about 128,000 micro-dramas, of which roughly 122,000 were AI micro-dramas, a share exceeding 95%. In the same period, research-institution data shows that among Douyin's monthly TOP 5000 short dramas, the number of fully AI-generated micro-dramas grew from 4 in January 2025 to 69 in October and 217 in November.

The direct consequence of this penetration rate is that the variance in quality has been amplified to the extreme. On the cost side, a mid-tier live-action short drama of 60 episodes at 2 minutes each costs roughly 500,000–600,000 CNY; an AI short drama of the same spec costs about 2,000–3,000 CNY per episode and under 200,000 CNY for the whole series (practitioner-interview basis, as cited in People's Daily, 2026). The flip side of an order-of-magnitude cost reduction is that many works fail to meet standards on "story coherence" and "character consistency."

The 21st Century Business Herald (2025-10-23) summary of industry problems can serve as this direction's "common failures" baseline:

• Content layer: no phenomenon-level hit has yet emerged; most rely on platforms to carry them; excessive pursuit of visual splendor and smooth motion while neglecting story coherence and character building, devolving into a "slide-show pile-up"; severe homogenization, clumping around xuanhuan (fantasy) and sci-fi themes.
• Tech layer: insufficient emotional expression and character-portrayal capability; character-and-scene consistency problems occur frequently.
• Commercial layer: overall profitability is not yet fully validated; most rely on platform subsidies and traffic support; weak IP attributes and low user retention.
• Copyright and ethics: some production crews use AI to generate scripts in one click, deepfake celebrities, and even sell "customized intimate scenes."

1.2 Definition and Scope

In this direction, AI Harness refers to the engineered carrier layer that, in AI micro-drama / online-drama production scenarios, takes on the five-stage pipeline of script → storyboard → generation → editing → dubbing, organizing scattered video-generation capabilities into a controllable, reproducible, regression-verifiable finished-production system.

1.3 Position within the AI Harness System

图 1-1｜AI 网剧 Harness：六层体系与五段流水线

数据来源：基于本文分析绘制的示意图。

Dominant layer: L3 Orchestration and Control layer + L4 Memory and State layer. Bottleneck layer: L4 Memory and State layer.

This direction exhibits the L4 bottleneck most typically. There are three reasons:

1. The longest time span crossed: a 60-episode short drama's production cycle is measured in weeks; a model session cannot retain state that long.
2. The greatest number of consistency dimensions: face, apparel, props, scene, and vocal timbre—five dimensions all must stay consistent; a drift in any one of them breaks immersion.
3. High regeneration cost: when character drift is discovered at episode 40, regenerating means rerunning the entire pipeline.

Therefore, the core structure of L4 in this direction is not "memory," but the shot state machine: recording generation state, version, verification results, and failure reasons on a per-shot basis, so that any single shot can be regenerated independently without affecting other shots.

1.4 Industry Status and Published Quantitative Metrics

Video-generation model specifications (consolidated from securities research reports, Haitong, 2025-10-29):

The spec differences directly determine the orchestration strategy: single-generation length is limited, which necessarily requires a "segmented generation + consistency checking + regeneration" pipeline structure — this is the physical premise of this direction's L3 orchestration-layer design.

2. Glossary

3. Case Studies

3.1 Kunlun Tech SkyReels: A Full-Pipeline Platform for AI Short Dramas

3.1.1 Background

In August 2024, Kunlun Tech launched SkyReels, positioned as the world's first AI short-drama creation platform integrating a video foundation model with a 3D foundation model. The core problem it set out to solve: AI short-drama production involves six types of capabilities—script, character, storyboard, video, audio, and editing—spread across different tools, with all intermediate state transfer done manually, causing consistency to degrade at every handover.

3.1.2 Approach

SkyReels built a complete pipeline whose structure is itself a reference implementation of an AI online-drama Harness:

From a Harness perspective, SkyReels' most critical design is the coupling of the character library with storyboard generation: storyboard generation directly references assets in the character library rather than re-describing the character each time. This is precisely the engineering embodiment of turning L4 from "memory" into a "state carrier."

3.1.3 Results

• Scripts generated by SkyScript consistently achieve a human rating of A/S grade.
• Video-generation specifications reach 1080P / 60 fps / 180 seconds per generation.
• The Kunlun Tech SkyReels platform plus the DramaWave distribution platform reached US $10 million in monthly revenue as of Q1 2025.

Sources: Kunlun Tech official website; Securities Times, Summary of Kunlun Tech's 2024 Annual Report, 2025. The monthly revenue figure is the enterprise-disclosed basis.

3.2 Bai Hu and Shan Hai Qi Jing: Compressing Cycle Time and Cost

3.2.1 Background

The cycle time and cost structure of traditional micro-dramas are determined by three parts: crew labor, location and shooting, and post-production. For xuanhuan (fantasy), sci-fi, and costume genres in particular, location and post-production account for an especially high share—a director once said that 90% of one drama's scenes took place in Xing'anling, and "if we were to shoot on location in the forested snowfields, the cost would be extremely high and the danger factor very large."

3.2.2 Approach

• Bai Hu (a fully AI-produced micro-drama): a 4-person team, using ChatGPT for rapid script iteration + AI image generation + intelligent editing.
• Shan Hai Qi Jing: Ji Po Zhan Lang (Kuaishou, director Chen Kun, 2024): produced through an all-AI end-to-end process.
• Xing'anling Gui Shi (China's first paid AI short drama, director Ding Kuan): uses AI to take over "the most tedious, money-burning, and most dangerous work."

Breaking it down from a Harness perspective:

• L2: an extremely small team (4 people) means the toolchain must be highly integrated; any intermediate format requiring manual handling becomes a bottleneck.
• L3: the script → storyboard → generation → editing → dubbing pipeline must be highly automated in a 4-person team, otherwise cycle-time compression is out of the question.
• L4: small teams are more prone to mistakes—with no dedicated character-asset-management role, consistency risk is higher. This explains why AI short dramas "frequently suffer character-and-scene consistency problems."

3.2.3 Results

Sources: 21st Century Business Herald, 2025-10-23; public reports, 2025.

3.3 AI Short Drama Strategies of Listed Companies and Platform Ecosystems

3.3.1 Background

In 2025, AI short dramas moved from the "creator experiment" stage into the "institutionalized production" stage. Listed companies, platforms, and tool providers entered at the same time, forming a complete ecosystem of production—tools—distribution—incentives.

3.3.2 Approach

From a Harness perspective, the defining feature of this stage is the integration of tools and distribution: SkyReels paired with DramaWave, JiMeng with Douyin, and Kling with Kuaishou. Feedback data from the distribution side (completion rate, retention) flows back to the production side, forming an L5 evaluation loop—something impossible to achieve at the small-team stage.

3.3.3 Results

• Guomai Culture: 240 episodes, 70 million all-network views.
• ChineseAll: nearly a hundred AI comics and dynamic comics, with cumulative views exceeding 3 billion.
• Jingying Technology: overseas AI short drama topped the weekly chart, with heat value exceeding 5 million.
• Industry level: signals from Hengdian World Studios show that in 2023 dozens of scenes were customized for micro-drama crews, while in the same period of 2026 the number of crews coming for on-location shooting dropped sharply (People's Daily investigation).

Sources: 21st Century Business Herald, 2025-10-23; People's Daily, 2026.

4. Practice Standards

4.1 AGENTS.md Specification

4.1.1. AGENTS.md (AI Drama Direction)

# AGENTS.md —— AI 网剧（AI Drama）

## 角色与边界
- 你运行在 AI 短剧制作流水线之上，负责剧本、分镜、生成、剪辑、配音五段任务。
- 你负责执行与制作，不负责创意方向与内容价值判断。
- 你不得在无具名人类导演或制片人签核的情况下输出成片。
- 你不得生成任何真实自然人（明星、演员、公众人物、普通个人）的肖像与声音。
- 你不得虚假标注主创信息。

## 环境假设
- 存在角色资产库：每个角色有三视图（正/侧/背）、表情集、色卡、音色与细粒度属性标签。
- 存在剧本与分镜系统；存在视频生成、TTS、剪辑、BGM 工具。
- 存在非编与渲染农场，支持批量渲染与导出封装。
- 存在镜头状态机：以镜头为单位记录生成状态、版本、校验结果与失败原因。
- 存在前置审核流程（AI 生成内容必须前置审核，而非事后抽查）。
- 存在 scripts/ 目录承载确定性操作：抽帧、色彩校正、元数据写入、标识渲染、一致性比对、导出封装。

## 上下文加载顺序（Context Budget）
1. 用户显式指令与本次镜头的验收标准
2. 本文件（方向级）与组级 AGENTS.md
3. 角色资产：三视图、表情集、色卡、音色及其版本号（不可裁剪）
4. 当前镜头在镜头状态机中的上下文：前一镜头收尾状态、本镜头分镜描述（不可裁剪）
5. 剧本当前段落与人物小传
6. 风格参考集与历史素材（可裁剪）
规则：角色三视图与前一镜头状态属于锚定上下文，任何情况下不得被裁剪。角色漂移的根因几乎总是锚定信息被挤出上下文。

## 工具契约
- 只读类（自由调用）：查询角色库、查询剧本、查询分镜、查询镜头状态机、查询素材。
- 生成类（输出必须进校验关卡）：剧本生成、角色形象生成、分镜生成、视频生成、TTS、BGM 匹配。
- 校验类（阻断式）：一致性比对（面部/服饰/道具/场景/音色）、标识校验、版权与肖像核查、内容前置审核。
- 写操作类（二次确认 + 留痕）：覆盖资产库版本、提交成片、提交分发。
- 抽帧、色彩校正、元数据写入、标识渲染、一致性比对、导出封装一律走 scripts/。

## 任务执行流程（SOP）
1. 确认本次任务：集数、镜头号、分镜描述、时长与规格。
2. 加载角色资产与前一镜头状态，锁定版本号。
3. 生成：按镜头类型选择模型与参数；输出 2~3 条候选。
4. 一致性校验：与三视图做面部/服饰/道具比对；与前一镜头做场景与色调比对。
5. 重生成回路：不通过则调整锚定输入后重生成；**同一镜头连续 2 次不通过则升级人工**。
6. 音频：从角色音色库取音色，按情绪匹配 BGM。
7. 标识处理：显著标识写入；隐式元数据写入。
8. 前置审核：内容合规、版权、肖像、主创信息真实性。
9. 人工签核：导演或制片人确认。
10. 导出与回读：导出封装后回读，确认元数据与标识在位。
11. 更新镜头状态机：记录版本、校验结果、失败原因、成本。

## 验证与证据要求
- 每个镜头必须留存：候选素材路径、比对分数、通过/不通过判定、失败原因（若失败）。
- 跨集交付时必须提供角色一致性抽查报告（抽查比例与判定标准由制片人确定）。
- 标识必须给出证据：显著标识位置 + 隐式元数据回读结果。
- 主创信息必须真实：不得将 AI 生成标注为真人导演/演员作品。
- 引用效果数据必须标注口径层级。

## 失败与升级策略
- 角色漂移（面部/服饰/道具）：回滚到镜头状态机最近通过点，重新加载三视图后重生成；连续 2 次失败升级人工。
- 场景漂移：把前一镜头的收尾帧作为参考图注入。
- 音色漂移：强制从角色音色库取值，禁止临时生成。
- 情感表达不足或人物塑造单薄：不强行加特效，回到剧本与表演设计环节。
- 内容同质化：在剧本阶段增加题材与结构约束，而非在生成阶段补救。
- 前置审核不通过：不得进入分发环节，按审核意见修改后重审。
- 升级必须携带：集数、镜头号、状态机记录、比对证据、已尝试处理。

## 安全与合规红线
- 建立 AI 训练数据版权合规机制；训练与参考素材必须有授权台账。
- 生成内容需符合社会主义核心价值观，不得生成违法违规内容。
- 不得侵害他人知识产权、隐私权、名誉权；禁止 AI 换脸明星。
- AI 生成内容必须进行**显著标识**；并写入隐式元数据。
- **不得虚假标注主创信息**。
- 建立内容审核机制，**对 AI 生成内容进行前置审核**。
- 不得恶意删除、篡改、伪造、隐匿标识（第十条）；去标识日志留存不少于六个月（第九条）。

## 禁止事项
- 禁止使用真实自然人肖像与声音，包括换脸与音色克隆。
- 禁止虚假标注导演、编剧、演员等主创信息。
- 禁止跳过一致性校验与前置审核。
- 禁止把角色状态存放在模型会话里；必须写入镜头状态机。
- 禁止让模型逐 token 生成抽帧结果、调色参数、元数据与导出文件。
- 禁止编造规范编号、条款与案例数值。

## 输出格式
集数 / 镜头号 / 使用角色资产版本 / 分镜描述 / 候选数量与路径 / 一致性比对结果 / 标识证据 / 前置审核结果 / 待签核项 / 责任人 / 状态机更新记录 / 遗留问题

## 评估与自检
- 角色三视图是否进入了上下文且未被裁剪？
- 本镜头的状态是否已写入状态机（不依赖会话记忆）？
- 一致性比对分数是否留存？
- 标识是否可回读？
- 主创信息标注是否真实？
- 本次失败样本是否已纳入评估集？

4.2 SKILL.md Specification

4.2.1. SKILL.md (AI Drama Single-Shot Compliant Delivery)

---
name: ai-drama-shot-delivery
description: AI 网剧单镜头交付技能。当需要从分镜描述出发生成一个镜头素材，并完成角色一致性比对、场景与色调连续性校验、显著标识与隐式元数据写入、内容前置审核、导演签核与镜头状态机更新时使用。适用于 AI 微短剧、AI 网剧的分段批量生产场景。
version: 1.0
created: 2026-09-12
---

# AI 网剧 · 单镜头合规交付

## 适用场景
- 从分镜描述与角色资产出发生成一个镜头的可交付素材。
- 需要在分段生成模式下保证角色、服饰、道具、场景与音色的跨镜头一致性。
- 需要把每个镜头的生成状态、校验结果与失败原因写入镜头状态机，支持单独重生成。

## 前置条件
- 已加载方向级 AGENTS.md 与组级 AGENTS.md。
- 角色资产库可用：三视图（正/侧/背）、表情集、色卡、音色，且带版本号。
- 镜头状态机可用，能读到前一镜头的收尾状态与色调基线。
- scripts/ 中存在抽帧、色彩校正、一致性比对、元数据写入、标识渲染、导出封装脚本。
- 已确定一致性判定阈值与抽查比例，并经制片人确认。
- 已确定具名签核人（导演或制片人）与前置审核流程。

## 输入
| 输入项 | 说明 | 必需 |
|---|---|---|
| 集数与镜头号 | 用于状态机定位 | 是 |
| 分镜描述 | 景别、运镜、时长、表演要求 | 是 |
| 角色资产引用 | 角色 ID + 资产版本号 | 是 |
| 前一镜头状态 | 收尾帧、色调基线、场景 ID | 是 |
| 输出规格 | 分辨率、帧率、时长、画幅 | 是 |
| 一致性阈值 | 面部/服饰/道具/场景的判定阈值 | 是 |

## 输出
- 镜头素材（含 2~3 条候选）
- 一致性比对报告（面部/服饰/道具/场景/音色）
- 标识证据（显著标识位置 + 隐式元数据回读结果）
- 前置审核结果
- 签核记录与镜头状态机更新记录

## 执行步骤
1. 从镜头状态机读取本镜头任务与前一镜头收尾状态。
2. 加载角色三视图、表情集、色卡与音色，锁定版本号。
3. 按镜头类型选择模型与参数，生成 2~3 条候选。
4. 一致性比对：与三视图比对面部/服饰/道具；与前一镜头收尾帧比对场景与色调。
5. 重生成回路：不通过则调整锚定输入（补参考图、加约束条件）后重生成；连续 2 次不通过升级人工。
6. 音频处理：从角色音色库取音色，按情绪匹配 BGM。
7. 标识处理：脚本渲染显著标识并写入隐式元数据。
8. 前置审核：内容合规、版权、肖像、主创信息真实性。
9. 导演或制片人签核。
10. 导出封装并回读校验。
11. 更新镜头状态机：版本、比对分数、失败原因、成本、责任人。

## 质量标准（DoD）
- 面部/服饰/道具/场景/音色五项一致性判定全部通过，且留存比对分数。
- 显著标识在位；隐式元数据可回读；去标识日志留存不少于六个月。
- 前置审核通过并留痕。
- 主创信息标注真实，无虚假标注。
- 镜头状态机已更新，本镜头可单独重生成而不影响其他镜头。
- 引用效果数据标注口径层级（如"成本 50~60 万 → 不到 20 万"为从业者访谈口径）。

## 常见失败与处理
- 角色换脸：三视图未入上下文或被裁剪 → 重新加载并固化为不可裁剪区。
- 场景漂移：未引用前一镜头收尾帧 → 把收尾帧作为参考图注入。
- 帧间不一致（多出的轮子、错位的轴）：缩短单次生成长度，增加逐帧人工精修节点。
- 情感表达不足：不要靠加大特效掩盖，回到剧本与表演设计环节。
- 内容同质化（扎堆玄幻/科幻）：在剧本阶段增加题材约束与结构多样性要求。
- 音色跨集漂移：强制从角色音色库取值，禁止临时生成新音色。
- 沦为"幻灯片堆砌"：把叙事连贯性纳入判定标准，限制单镜头华丽度权重。
- 前置审核不通过：按审核意见修改后重审，不得绕过。

## 示例
任务：第 7 集第 12 镜头，中景，角色 A 在雨中回身，时长 6 秒，1080P / 30 fps / 竖屏。
输入：角色 A 资产版本 v3.2；前一镜头（SC-0711）收尾帧与色调基线；分镜描述；一致性阈值。
执行：加载三视图 → 生成 3 条候选 → 比对（候选 1、2 服饰偏差超阈值，候选 3 通过）→ 音色库取音色 → 标识渲染 → 前置审核 → 导演签核 → 导出回读 → 更新状态机。
输出：成片素材 + 比对报告（含 2 条失败样本）+ 标识证据 + 签核记录 + 状态机记录（SC-0712 = passed）。

4.3 Implementation Checklist

5. Summary

The AI online-drama direction exposes the Creative Industries cluster's L4 bottleneck most thoroughly. The reason is simple: it is the only direction that must keep "the same character" consistent across dozens of episodes, hundreds of shots, and a weeks-long production cycle. If any of the five dimensions—face, apparel, props, scene, vocal timbre—drifts, the audience is immediately thrown out of the story.

The industry's answer is not "a better model," but "a better state carrier." In Kunlun Tech's SkyReels design, the most valuable reference is not the 1080P / 60 fps / 180-second spec, but the coupling of the character library with storyboard generation—the storyboard directly references character-library assets rather than re-describing the character each time. This design turns consistency from a "model-capability problem" into an "engineering-structure problem."

The cost and cycle-time data are real: a whole AI short drama dropped from 500,000–600,000 CNY to under 200,000 CNY, Bai Hu compressed from 3 months to 2 weeks, and Shan Hai Qi Jing compressed from 3–6 months to 2 months with cost falling to 1/4. But one must also see the problems the industry has not yet solved: no phenomenon-level hit has yet emerged, severe homogenization, overall profitability not fully validated, and "frequently occurring character-and-scene consistency problems."

On the regulatory side, the China Netcasting Services Association released the Network Audio-Visual AI Short Drama Creation, Production, and Distribution Specification in April 2025, requiring conspicuous labeling of AI-generated content, no false credit attribution, establishment of a training-data copyright compliance mechanism, and pre-release review of AI-generated content. It must be stated truthfully: the official document number and full text of this specification could not be verified on the association's official website; this document cites it only from secondary sources, marked as [To be verified].

Information Gap Statement

6. References

1. Procedures for the Labeling of AI-Generated and Synthetic 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
2. AI Short Dramas: The New Hot Spot Being Chased by Capital — 21st Century Business Herald, 2025-10-23. https://www.21jingji.com/article/20251023/herald/89aa15c81c1bf4ae8b48bf94060ed94e.html
3. Will AI "Shuffle" Micro-dramas? — People's Daily, 2026. https://kpzg.people.com.cn/n1/2026/0511/c404214-40717117.html
4. Will AI "Shuffle" Micro-dramas? (Hunan Channel reprint) — People's Daily Online, 2026. https://hn.people.com.cn/BIG5/n2/2026/0511/c208814-41577065.html
5. AI Short Drama Specification Published (China Netcasting Services Association) — AIWW. Note: secondary source; the official document number has not been verified. https://www.aiww.cn/s/ff3ac3ba84deb97b
6. Summary of Kunlun Tech's 2024 Annual Report (SkyReels technology stack) — Securities Times, 2025. https://epaper.stcn.com/pic/202504/26/8eec6686f3528439d13a5db995a8f55b.pdf
7. Kunlun Tech Official Website (SkyReels-A1 and other open-source models) — Kunlun Tech. https://www.kunlun.com/en
8. Promoting the Construction of the Labeling System Through Multiple Measures to Support the Healthy Development of AI in the New Era (labeling implementation methodology and 6 standard practice guides) — National Internet Emergency Center, 2025. https://www.cac.gov.cn/2025-09/06/c_1758880709361356.htm
9. Research Report on the Development of 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
10. AGENTS.md Official Website — Agentic AI Foundation (Linux Foundation). https://agents.md/
11. Agent Skills Specification — agentskills.io. https://agentskills.io/specification
12. Equipping agents for the real world with Agent Skills — Anthropic, 2025 (updated 2025-12-18). https://claude.com/blog/equipping-agents-for-the-real-world-with-agent-skills