Most AI ad tools write copy. They do not know that a US skincare claim needs substantiation, that 绝对化用语 is illegal in Chinese advertising, that an AI-generated creator must not deliver first-person usage testimony, or that a storyboard shot and its voiceover line have to come from the same source of truth or the shoot falls apart.
tiktok-ugc-ad-maker is an Agent Skill that treats an ad plan as a production document with a quality gate, not a text generation. Every downstream asset — spoken script, B-roll list, video-generation prompt — is derived from one storyboard JSON, and the plan does not ship until a consistency checklist passes.
A single run produces a 14-section production package:
| # | Section | What it is |
|---|---|---|
| 0 | 小白解释 | Only the terms this task needs, with why each matters for conversion |
| 1 | 制作参数确认 | Country, platform, organic vs paid, TikTok Shop card, language, voice format, duration, aspect ratio |
| 2 | 默认假设 | Every default chosen for you, stated openly and editable |
| 3 | 产品理解 | Features translated into physical / emotional / routine / proof hooks |
| 4 | 广告结构 | Timeline with 时间 · 内容 · 目的, hook inside the first 2s |
| 5 | 合规提示 | Disclosure, AIGC labelling, music rights, claim boundaries |
| 6 | 达人设定 | Fixed creator reference or designed virtual creator |
| 7 | Hook 库 | Exactly 10 hooks, A/B ready |
| 8 | 分镜表 + JSON | The single source of truth for everything downstream |
| 9-11 | 口播 / B-roll / 视频提示词 | Each block maps back to a storyboard shot ID |
| 12 | 分镜一致性核对表 | Must pass before delivery |
| 13 | 发布前检查 + 复盘 | Pre-flight checklist and post-launch iteration |
Load the repository in any agent runtime that supports local skills (Claude Code, Codex, or your own harness):
git clone https://github.com/ROTl24/tiktok-ai-skills.gitPoint your agent at tiktok-ugc-ad-maker/, then upload a product image and ask:
用 tiktok-ugc-ad-maker 给这个产品做一版美国 TikTok 15 秒真人口播广告方案
The skill confirms six parameters before writing anything:
| Parameter | Options |
|---|---|
| 投放国家 / 地区 | US TikTok, 中国抖音, other markets |
| 发布平台与投放方式 | 自然种草 / 付费投放, TikTok Shop 挂车 是/否 |
| 口播语言 | any |
| 声音形式 | 真人口播 / 旁白 |
| 达人 | uploaded reference (@Image1) or designed virtual creator |
| 视频时长 | 15s / 30s / 45s |
Say "你来决定" and it states its assumptions explicitly instead of hiding them.
The storyboard is not a deliverable among others — it is the origin. Section 12 fails the run if any shot ID is missing downstream, which is what stops the classic failure mode where the voiceover describes a shot the storyboard never planned.
Coverage depth is deliberately asymmetric and stated as such — an honest compliance surface beats a uniform-looking one:
| Market | Depth | Covers |
|---|---|---|
| 🇺🇸 US TikTok | Deepest | Beauty/skincare claims, FTC disclosure, FDA cosmetic claim boundaries, influencer rules, AI-creator disclosure |
| 🇨🇳 中国抖音 | Baseline | 广告法绝对化用语, 品类规则, 价格促销, AIGC 标识 |
| 🌍 Other markets | Generic | General rules, with an explicit prompt to verify local platform policy |
Compliance references carry a last-verified date in their header. If the date is stale, or the category is restricted, the skill tells you to check current official policy before drafting direct-response copy.
These are the rules the skill will not negotiate on:
- Refuses to advertise products that appear prohibited, counterfeit, unsafe, or too regulated to assess — and explains the blocker instead of producing a softened ad.
- No fabricated testimony from virtual creators. An AI-designed creator never says "I've been using this for months." It uses observable demonstration and in-frame reaction, and the AIGC label note is mandatory.
- No invented substance. Ingredients, certifications, test results, and prices are never generated.
- No placeholder output. "Show product", "highlight benefits", "add CTA" fail the quality gate — every section must carry product-specific detail drawn from the actual image.
- Reference discipline.
@Image1creator,@Image2scene/style,@Image3product; each shot declares only the references it actually needs.
| Mode | Use when |
|---|---|
| Full package | You want the complete 14-section document in one pass |
| Staged delivery | Confirm the storyboard first, then generate script / B-roll / prompts |
| Incremental update | Change one shot or hook; downstream modules update and the checklist re-runs |
| Hook variant | Swap only the opening for A/B testing, reuse everything else |
tiktok-ai-skills/
├── tiktok-ugc-ad-maker/
│ ├── SKILL.md # workflow, quality gate, output contract
│ ├── agents/openai.yaml
│ └── references/
│ ├── general-tiktok-product-compliance.md
│ ├── us-tiktok-beauty-compliance.md
│ ├── cn-douyin-compliance.md
│ ├── output-quality-bar.md
│ ├── example-output.md # gold-standard sample
│ └── iteration-playbook.md
└── docs/test-results/ # dated verification records
references/example-output.md is the quality baseline. Structural changes to SKILL.md require updating the sample and leaving a verification record in docs/test-results/.
- Output serves a real shoot and a real ad buy, not a demo.
- Prefer specific product information over generic templates.
- Section numbering and output structure are defined once in
SKILL.md; other documents reference by name to avoid sync drift. - Compliance notes exist to keep an ad executable — they are not a substitute for legal review.
- Deeper coverage for EU / UK / SEA markets
- Additional skills in the repository beyond
tiktok-ugc-ad-maker - Machine-readable compliance rule files
- More gold-standard samples across product categories
Issues and PRs welcome. When changing skill behaviour:
- Update
SKILL.md— it is the single definition of structure. - Regenerate
references/example-output.mdso the sample matches. - Add a dated verification record under
docs/test-results/.
MIT © 陆振豪 (ROTl24)
tiktok-ugc-ad-maker 是一个 Agent Skill,把产品图或产品说明转成可直接开拍的中文 UGC 广告方案。
它和普通 AI 文案工具的区别在于两点:
一是合规前置。 在写任何 hook 和卖点之前,先做品类与风险分级,再按目标市场路由到对应合规文档。美国 TikTok 覆盖最深(美妆专项、FTC 披露、FDA 化妆品宣称边界),中国抖音提供基础参考(绝对化用语、价格促销、AIGC 标识)。覆盖深度不对称这件事是明写出来的,不装作全都很懂。
二是分镜作为唯一事实源。 口播稿、B-roll 清单、视频生成提示词全部从同一份分镜 JSON 派生,每个时间块都要映射回分镜 shot ID;第 12 节的一致性核对表不通过就不交付。这挡掉了最常见的翻车方式——口播在描述一个分镜里根本没有的镜头。
护栏部分不可协商:违禁 / 假冒 / 无法评估的商品直接拒绝出稿;AI 虚拟达人不得编造使用史证言,只能用可观察的演示和镜头内即时反应,并强制 AIGC 标注;成分、认证、检测结果、价格一律不得虚构。
实际投放前仍需自行核对目标市场政策、落地页信息与广告主资质。