AI Harness in Media

1. Introduction

1.1 Background

Media is the direction where AI Harness faces the most rigid compliance constraints and the most sensitive responsibility structure upon deployment. There are three reasons:

First, the dual identity. A media organization is both the generator of synthetic content and the disseminator facing the public, so it bears both the labeling obligation on the generation side and the verification obligation on the dissemination side. Article 6 of the Measures for Labeling AI-Generated Synthetic Content (CAC Letter No. 2 of 2025, effective 1 September 2025) imposes four obligations on dissemination platforms: verify the implicit label in the metadata and then add a prominent notice; where there is no implicit label but the user declares content as AI-generated, so indicate with "may be"; where there is no label but generation traces exist, identify it as "suspected" and so indicate; and provide a labeling function and remind users to declare. These four obligations must be implemented one by one into Harness's L6 governance layer, rather than left to editorial discretion.

Second, the asymmetrical cost of factual errors. When a single factual error appears in an article, what is lost is the organization's credibility, and its impact far outweighs the gain from a single improvement in production efficiency. Therefore, the core of L1 context engineering in the media direction is not "let the model see more" but "let the model see only trusted sources" — source grounding is the first principle of L1 in this direction.

Third, scaling has already happened. According to CCTV.com's September 2025 report, more than 490 large models in China have completed filing with the Cyberspace Administration of China (CAC), more than 240 have completed registration with provincial CAC offices, and the user base of generative AI products has reached 230 million people. At this scale, "manual item-by-item review" is no longer a sustainable quality strategy — verification and labeling must be turned into pipeline checkpoints.

1.2 Definition and Scope

In this direction, AI Harness refers to the engineered bearer layer that, in the scenarios of news production, content distribution, and converged media operations, takes on the full-chain tasks of planning, collection, production, review, distribution, and feedback, turning model output into publishable, traceable, and verifiable content assets.

1.3 Position within the AI Harness System

图 1-1｜媒体方向 AI Harness 六层定位与瓶颈分布

数据来源：基于本文分析绘制的示意图。

Leading layer: L1 Context Engineering layer + L6 Governance and Security layer. Bottleneck layer: L4 Memory and State layer + L6 Governance and Security layer.

Consistent with Creative / AI web series / AI animation, the quality bottleneck in the media direction lies in L4. Concretely: institutional tone, terminology tables, expression norms, and layout style need to remain consistent across journalists, departments, and time — yet such knowledge usually exists only in the minds of senior editors. Making it explicit as a "loadable house style context" is the most valuable and yet most easily skipped step in building Harness for the media direction.

At the same time, the L6 bottleneck of the media direction is distinctive: it is not only about "what must not be done" but also about "what must be done" — labels must be applied, verification must be performed, notice must be given. This is a set of positive obligations, different in nature from the "prohibitive red lines" of the industry direction; in engineering terms it requires "validation + blocking" rather than mere "interception".

1.4 Industry Status and Published Quantitative Baselines

2. Glossary

3. Case Studies

3.1 Xinhua's Converged Production Engine and the "Editorial Assistant"

3.1.1 Background

A news agency's production is characterized by "high concurrency + high standards + multilingual + multi-format": the same event must produce multiple versions — text, images, video, social media, and more — within an extremely short time, and any single factual error gets amplified by global media. In a traditional production system, each stage of planning, collection, editing, and review is hosted by a different system; data and context are not shared, and model capabilities struggle to enter the mainline.

3.1.2 Approach

Based on its mainstream value corpus and a media-domain large model, Xinhua built a converged production engine; the "editorial assistant" shown at the 2025 China New Media Technology Exhibition integrates 227 functions including content verification, trend discovery, and multimodal generation. In addition, according to Journalism and Writing Issue 4, 2025, Xinhua independently built the "controllable, reliable, multi-agent converged production engine Xinhua Xinyu", which stands out in content verification, false-information identification, intelligent situation analysis, knowledge-enhanced intelligent planning, and vision-enhanced intelligent creation.

Deconstruction from a Harness perspective:

• L1: The mainstream value corpus constitutes a strongly constrained context, with source grounding built directly into the engine.
• L2: The 227 functions are essentially 227 registered tools, forming a complete media toolset.
• L3: The multi-agent converged production engine is the productized form of the L3 orchestration layer.
• L5: Placing "content verification" and "false-information identification" on par with "intelligent creation" shows that evaluation and verification are built-in links of the production chain rather than add-ons.

3.1.3 Results

• The editorial assistant integrates 227 functions, covering content verification, trend discovery, multimodal generation, and more.
• A controllable, reliable production engine based on the mainstream value corpus has been formed.

Source: Xinhua News Agency, 2025 (2025 China New Media Technology Exhibition); Journalism and Writing Issue 4, 2025.

3.2 Hunan Broadcasting Group's "Mango Large Model" Full-Process Enablement

3.2.1 Background

A broadcast organization's content production chain is longer than a news agency's: from IP planning, scripts, filming, post-production, and packaging to broadcast and operations, there are more links, more roles, and fragmented systems. Single-point AI tools (such as auto-generated subtitles) can only improve local efficiency — they cannot change the overall production model.

3.2.2 Approach

Hunan Broadcasting Group implemented an "AI + Audio-Visual" initiative and independently developed the "Mango Large Model" for full-process AI enablement. According to public information, Mango Excellent Media's "Shanhai Research Institute" and the "Mango Large Model" have passed CAC filing.

Deconstruction from a Harness perspective:

• L1: Centered on vertical-domain corpora across variety shows, TV dramas, animation, sports, and more (consistent with the approach of CMG's CCTV-Ting Large Model 2.0 launched in February 2025, both centering on vertical-domain corpora).
• L3: Full-process enablement means the orchestration covers the complete chain from planning to operations, rather than single-point tools.
• L4: Persistent management of vertical-domain corpora and existing IP assets is the foundation for consistency across programs and quarters.
• L6: Passing CAC filing satisfies the compliance prerequisite for generative AI services.

3.2.3 Results

• Business application coverage exceeds 93%.
• Over 83 innovative scenarios deployed.
• Over 180,000 videos generated cumulatively.

Source: Xinhua, 2025. The figures are data publicly presented by the organization at an industry conference.

3.3 Generative AI Podcast: NotebookLM's Deployment in the Newsroom

3.3.1 Background

Podcasting is one of the lowest return-on-effort formats for media organizations: a ten-minute episode typically takes hours from topic selection, writing, recording, to post-production. Nigeria's Premium Times and Vietnam's Nhan Dan (People's Daily) each tried using NotebookLM's Audio Overview capability in 2025 to automatically turn articles into podcasts.

3.3.2 Approach

The workflow: journalist imports the article → adjusts the prompt → generates the podcast → multi-platform distribution.

On the Vietnamese side, NotebookLM supported Vietnamese at the end of April 2025; the AI podcast released by editor-in-chief Le Quoc Minh was commented by listeners as "sounding like a real person". ONECMS product director Bui Cong Duyen measured: using NotebookLM to produce a podcast about "Resolution No. 68" took 5 minutes.

Practitioner's boundary judgment (Ha Thuy): AI suits simple podcasts such as news, guides, and book summaries; complex podcasts — roundtables, personal sharing, and emotional reflection — require human-machine collaboration plus full post-production review.

Deconstruction from a Harness perspective:

• L2: Generation is a tool call, but distribution is a write operation requiring secondary confirmation.
• L3: Import → prompt tuning → generation → human review → distribution; the review node sits between generation and distribution and cannot be omitted.
• L5: The human review rate is the key observable in this scenario.

3.3.3 Results and Constraints

Results:

• Production time went from 1~2 hours down to a few minutes (simple type); a single-item test took 5 minutes.
• Premium Times feedback: "Production became quicker, significantly reducing time in podcast creation."

Constraints and known limitations (the core value of this case is that it is an honest failure list):

Source: INMA (International News Media Association), 2025; Vietnam.vn, 2025.

4. Practice Standards

4.1 AGENTS.md Specification

4.1.1. AGENTS.md (Media · Media direction)

# AGENTS.md —— 媒体（Media）

## 角色与边界
- 你运行在媒体融合生产平台之上，负责策划辅助、稿件与多媒体生产、事实核查、多平台分发建议四类任务。
- 你是助手，不是记者，也不是编辑。你不得替人类做价值判断与选题决策。
- 你不得在无具名责编签核的情况下发布任何内容。
- 你对事实的陈述必须能指到具体信源；无法指到信源的内容一律不得输出。

## 环境假设
- 存在主流价值语料库与机构风格规范（调性、术语表、禁用表述），可加载为上下文。
- 存在采编系统、非编工具、字幕与语音合成、智能审核、媒资管理系统。
- 存在信源白名单与事实核查工具。
- 存在分发平台 API（写操作，需二次确认）。
- 存在 scripts/ 目录承载确定性操作：元数据写入、标识渲染、抽帧、转码封装、标识回读校验。

## 上下文加载顺序（Context Budget）
1. 用户显式指令与本次任务的事实边界
2. 本文件（方向级）与组级 AGENTS.md
3. 机构风格规范与术语表（不可裁剪）
4. 信源白名单与已核实事实库（不可裁剪）
5. 相关选题库条目、媒资库素材、已发布内容台账
6. 检索结果与参考素材（可裁剪）
规则：信源白名单与机构风格规范属于锚定上下文，任何情况下不得被裁剪。

## 工具契约
- 只读类（自由调用）：检索语料、查询媒资、查询已发布台账、调用事实核查。
- 生成类（输出必须进校验关卡）：撰稿、改写、摘要、文生图、语音合成、视频剪辑、播客生成。
- 合规类（阻断式）：标识写入、标识回读校验、肖像与版权核查、敏感题材识别。
- 写操作类（二次确认 + 留痕）：发布、撤稿、修改已发布内容、调用分发 API。
- 元数据写入、角标渲染、抽帧、转码封装一律走 scripts/，禁止模型逐 token 生成。

## 任务执行流程（SOP）
1. 确认任务类型（策划 / 生产 / 核查 / 分发）与事实边界。
2. 加载机构风格规范与信源白名单，锁定版本。
3. 取素材：只使用白名单内信源与已授权素材。
4. 生成：产出候选稿或候选素材。
5. 事实核查：逐条指到信源；无法指源的删除或标注待核实。
6. 标识处理：生成合成内容必须写入显式标识与隐式元数据。
7. 合规校验：敏感题材、肖像权、版权、标识完整性。
8. 人工签核：具名责编确认。
9. 发布与回读：发布后回读，确认标识在位、元数据完整。
10. 归档：写入已发布内容台账与媒资库，记录版本、信源、轨迹、责任人。

## 验证与证据要求
- 每个事实陈述必须附信源：机构 + 时间 + 链接或文献位置。
- AI 生成或辅助生成的内容必须可追溯：模型、版本、提示摘要、生成时间。
- 标识必须给出证据：显式标识的位置截图或渲染结果；隐式元数据的回读结果。
- 引用效果数据必须标注口径层级。
- 无法核实的内容一律标注 [待核实]，不得模糊化处理。

## 失败与升级策略
- 事实无法核实：删除该陈述或标注 [待核实]，升级到责编。
- 信源冲突：并列呈现冲突信源，不得自行选择，升级到责编。
- 标识缺失或回读失败：不得发布，回到导出环节重做。
- 敏感题材（涉政、涉民族宗教、涉未成年人、涉医疗健康）：停止自动流程，转人工审核。
- 传播端发现疑似未标识的 AI 内容：按《标识办法》第六条提示"疑似"，不得直接删除或放行。
- 升级必须携带：任务 ID、阶段、证据路径、已尝试处理。

## 安全与合规红线
- 严格遵循《人工智能生成合成内容标识办法》（国信办通字〔2025〕2 号，2025-09-01 施行）：生成端添加显式标识与隐式标识；传播端核验元数据并加显著提示；无隐式标识但用户声明的提示"可能为"；无标识但有生成痕迹的识别为"疑似"并提示。
- 不得恶意删除、篡改、伪造、隐匿标识，不得为他人实施上述行为提供工具或服务（第十条）。
- 去标识相关日志记录留存不少于六个月（第九条）。
- 不得侵害他人知识产权、隐私权、名誉权；不得使用未授权肖像与声音。
- 不得虚假标注主创信息。

## 禁止事项
- 禁止在无信源的情况下陈述事实。
- 禁止跳过事实核查与人工签核环节。
- 禁止发布未打标识或标识回读失败的生成合成内容。
- 禁止使用未授权的图片、音频、视频与肖像。
- 禁止把机构风格规范视为可裁剪上下文。
- 禁止编造信源、数据、引语、标准编号。

## 输出格式
任务编号 / 任务类型 / 使用信源清单 / 机构风格版本 / 内容产物路径 / 事实核查记录 / 标识证据 / 合规校验结果 / 待签核项 / 责任人 / 遗留问题

## 评估与自检
- 每个事实是否都能指到信源？
- 生成内容是否都打了标识且回读成功？
- 是否使用了白名单外的素材？
- 人工复核率是否异常偏高（说明系统不可信）？
- 本次任务是否已纳入评估集（含失败样本）？

4.2 SKILL.md Specification

4.2.1. SKILL.md (Media · a compliant AI-assisted content delivery)

---
name: media-compliant-content-delivery
description: 媒体方向内容交付技能。当需要完成一条 AI 辅助生成的新闻稿件、短视频、音频或播客，并必须走完信源核查、事实核查、AIGC 显式与隐式标识写入、传播端核验、责编签核与归档全流程时使用。适用于通讯社、广电、报社与融媒体中心的生产场景。
version: 1.0
created: 2026-09-12
---

# 媒体 · 一条合规的 AI 辅助内容交付

## 适用场景
- 从选题或素材出发，产出一条可发布的新闻稿件、短视频、音频或播客。
- 内容中部分或全部由 AI 生成，需要完成 AIGC 标识义务。
- 需要留下可追溯的信源、核查、标识与签核证据链。

## 前置条件
- 已加载方向级 AGENTS.md 与组级 AGENTS.md。
- 已加载机构风格规范（调性、术语表、禁用表述）与版本号。
- 信源白名单与已核实事实库可用。
- scripts/ 中存在元数据写入、标识渲染、转码封装与回读校验脚本。
- 已确定具名责编与各签核节点。

## 输入
| 输入项 | 说明 | 必需 |
|---|---|---|
| 选题或素材 | 事件描述、原始素材、素材授权状态 | 是 |
| 信源清单 | 可引用的信源及可信度分级 | 是 |
| 输出形态 | 稿件 / 短视频 / 音频 / 播客，及平台规格 | 是 |
| AI 参与范围 | 哪些环节由 AI 生成（决定是否触发标识义务） | 是 |
| 发布渠道 | 渠道清单与各自规格要求 | 是 |
| 截止时间 | 交付时间 | 否 |

## 输出
- 内容产物（稿件 / 成片 / 音频）
- 信源清单与事实核查记录
- 标识证据（显式标识位置 + 隐式元数据回读结果）
- 合规校验结果（肖像、版权、敏感题材）
- 签核记录与归档记录

## 执行步骤
1. 明确 AI 参与范围：只要内容含生成合成成分，即触发标识义务，进入标识流程。
2. 加载机构风格规范与信源白名单，锁定版本。
3. 收集素材：只使用白名单内信源与已授权素材；未授权素材停止使用并上报。
4. 生成候选：产出稿件或音视频初版。
5. 事实核查：逐条指到信源。无法指源的删除或标注 [待核实]；信源冲突的并列呈现，不自行选择。
6. 风格校验：对照术语表与禁用表述表逐项检查。
7. 标识处理：用脚本写入隐式元数据（内容属性信息、服务提供者名称或编码、内容编号），并在文本/图片/音频/视频的适当位置渲染显著提示标识。
8. 合规校验：肖像权、版权、敏感题材、主创信息真实性。
9. 责编签核：提交产物 + 全部证据，责编确认。
10. 发布与回读：发布后自动回读，确认标识在位、元数据未因转码丢失。
11. 传播端核验：若作为传播平台，按第六条核验上游隐式标识并加显著提示。
12. 归档：写入已发布内容台账与媒资库，记录版本、信源、轨迹、责任人。

## 质量标准（DoD）
- 每个事实陈述可指到具体信源（机构 + 时间 + 位置）。
- 显式标识在位且醒目；隐式元数据字段完整且可回读；去标识日志留存不少于六个月。
- 无未授权素材、肖像与声音；主创信息标注真实。
- 敏感题材已转人工审核并留痕。
- 产物有版本号；生成轨迹完整；责任人明确。
- 引用效果数据标注口径层级。

## 常见失败与处理
- 事实无法核实：删除陈述或标注 [待核实]，升级责编；不得模糊化措辞蒙混。
- AI 口音与本地化不符（如西方化口音）：改用本地化语音方案，或转为"AI 辅助 + 真人配音"。
- 可定制性不足（播客长度、语调、主持人数不可控）：把可控项前置到提示设计，不可控项在后期人工补齐。
- 转码后隐式元数据丢失：在封装脚本中重新写入，并对每个衍生版本回读校验。
- 组织采纳阻力：把"文化转型"纳入实施计划，先在小范围试点再推广。
- 复杂题材（圆桌、情感倾诉）AI 生成质量不足：改为人机结合 + 完整后期复核，不强行全自动。

## 示例
任务：把一篇已发布深度报道转成 8 分钟播客并分发到三个音频平台。
执行：导入原文 → 提示设计（长度与语调约束）→ 生成 → 事实核查（逐条指源）→ 脚本写入隐式元数据并在开头播报显著提示 → 责编签核 → 分发 → 回读校验。
结果：制作时间由 1~2 小时缩短至约 5 分钟（简单类口径）；标识证据与签核记录归档。
注意：圆桌讨论、个人分享类复杂播客不适用此全自动流程，须改为人机结合。

4.3 Deployment Checklist

5. Summary

The AI Harness for the media direction can be summarized in one sentence: make credibility an engineerable attribute.

On the capability side, scaling has already happened: Xinhua's editorial assistant integrates 227 functions, the Mango Large Model covers more than 93% of business, deploys over 83 scenarios, and has cumulatively generated more than 180,000 videos, and People's Daily Online's mainstream value corpus exceeds 300 TB. These numbers show that media organizations' AI applications have moved past the "pilot" stage.

But on the engineering side, the real difficulty lies in two places. The first is L4: the knowledge that "exists only in the minds of senior editors" — institutional tone, terminology tables, expression norms — must be made explicit as loadable context assets, otherwise every session is a fresh style drift. The second is L6: the Labeling Measures impose a set of positive obligations on media organizations — labels must be applied, verification must be performed, notice must be given, and logs must be retained for no less than six months — which differs in nature from the industry direction's "prohibitive red lines" and requires "validation + blocking" rather than mere "interception".

The value of the NotebookLM podcast case lies in the honest failure list it provides: limited customizability, westernized accents, mandatory human review, and the fact that organizational adoption is itself a cultural transformation. These constraints are not the flaws of any single tool but the general boundaries of current generative AI in professional content production. Writing the boundaries into Harness matters more than writing in the capabilities.

Information Gap Statement

6. References

1. Measures for Labeling AI-Generated 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. Advancing Labeling System Construction Through Multiple Measures for 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
3. Measures for Labeling AI-Generated Synthetic Content Helping Distinguish False Information — gov.cn, 2025. https://www.gov.cn/zhengce/202503/content_7014404.htm
4. From 1 September, AI-Generated Synthetic Content Must Be Labeled — CCTV.com, 2025. https://big5.cctv.com/gate/big5/news.cctv.cn/2025/09/01/ARTI3ZlXK7MyM39Pm3PuZ5Hm250901.shtml
5. Embracing the Intelligent Wave, Swimming Toward the Depths of Change (Xinhua editorial assistant, Mango Large Model, People's Daily Online corpus) — Xinhua, 2025. https://www.news.cn/20251114/1c01598d836c449bbfd54267b6ecea6d/c.html
6. Research Report on the Development of New Media by Mainstream Outlets (2024-2025) — People's Daily Online, 2025. https://sc.people.com.cn/BIG5/n2/2025/1030/c345167-41396739.html
7. Generative AI Opens a New Era in Media Development — Reading of the CMG AI Development White Paper (2025 Edition) — Modern Television Technology, 2025. https://tvtech.cctv.com/tuijian20250707/index.html
8. Premium Times adopts Google Notebook LM to streamline newsroom processes — INMA, 2025. https://www.inma.org/blogs/conference/post.cfm/premium-times-adopts-google-notebook-lm-to-streamline-newsroom-processes
9. The secret behind Vietnamese podcasts that surprised the online community — Vietnam.vn, 2025. https://www.vietnam.vn/en/bi-mat-phia-sau-podcast-tieng-viet-khien-cong-dong-mang-ngo-ngang
10. Assessing NotebookLM's Ability to Generate Podcasts for Patient Education on Cataracts — IOVS / ARVO, 2025. https://iovs.arvojournals.org/article.aspx?articleid=2807446
11. Official site of AGENTS.md — Agentic AI Foundation (Linux Foundation). https://agents.md/
12. Agent Skills Specification — agentskills.io. https://agentskills.io/specification
13. 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