Self Media Compliance Review

Self-Media Compliance Review

Core Rule

Run a compliance review before any self-media video, clip, cover, title, subtitle, voiceover, product link, or publish copy is treated as final. The review is a risk-control pass, not legal advice and not a guarantee that a platform will approve the content.

Default output language is Chinese unless the user asks otherwise.

Inputs

Collect or infer these inputs before judging:

• Target platform(s): 视频号, 微信公众号, 订阅号, 服务号, 抖音, 快手, B站, 小红书, TikTok, 抖音电商, etc.
• Public-facing material: final video path, script, subtitles, cover text, title, description, tags, comments, product link copy, account/profile copy.
• Context: account identity, topic vertical, target audience, source ownership/authorization, product/service being promoted, qualifications for regulated topics.
• Evidence access: timecodes, frame notes, transcript lines, copy files, screenshots, manifests, or links.

E-Commerce Extended Inputs

When the content involves 带货、挂车、商品链接、抖店、千川投放 or any commercial promotion, also collect:

• 商品详情页: product detail page URL, full-page screenshot, captured-at timestamp.
• SKU/规格: all SKU variants with price, weight, volume, color, size.
• 价格/赠品/活动规则: current selling price, original/line-through price, gift items with specs, activity deadlines, coupon/red-packet conditions.
• 资质与授权: business license, category-specific qualifications, brand authorization letter, patent/certification numbers.
• 素材来源: source files, license/authorization proof for video, images, BGM, fonts, portrait, and brand assets.
• 历史处罚: prior violations, rejected 千川素材, account penalty history.

If a required input is missing, continue with the available evidence and mark the item 待核验; do not invent facts, qualifications, authorizations, prices, source provenance, or platform behavior.

Platform References

Always run the universal audit areas below. Then load platform-specific references when relevant:

• 微信视频号 / 视频号 / WeChat Channels: read references/wechat-channels.md.
• 微信公众号 / 微信公众平台 / 订阅号 / 服务号 / WeChat Official Accounts: read references/wechat-official-account.md.
• 抖音 / Douyin: read references/douyin.md.
• 快手 / Kuaishou: read references/kuaishou.md.
• B站 / 哔哩哔哩 / Bilibili: read references/bilibili.md.
• 小红书 / Xiaohongshu / RED: read references/xiaohongshu.md.
• TikTok / 国际版抖音: no dedicated reference file yet — run the universal audit areas only.
• Recent enforcement cases, creator discussions, or account-status/限流 questions: read references/recent-cases-2025-2026.md.

E-Commerce References

When the content involves 带货、挂车、商品链接、抖店、千川投放、选品 or any e-commerce scenario, also load:

• 抖音电商规则: references/douyin-ecommerce.md.
• 千川低质素材治理: references/qianchuan-low-quality.md.
• 电商功效与声明规则: references/ecommerce-claims.md.
• 电商违规案例: references/cases/ecommerce-cases.md.

E-Commerce trigger words: 带货、挂车、商品链接、详情页、SKU、佣金、转化、成交、选品、千川、低质素材、抖店、精选联盟、商品三一致、功效承诺、前后对比、赠品、优惠券、到手价、小黄车、商品分享、直播带货、达人带货、商家、开店、罚.

When using recent cases, load the index first, then the matching platform case file under references/cases/:

• 小红书: references/cases/xiaohongshu.md.
• 微信公众号: references/cases/wechat-official-account.md.
• 视频号: references/cases/wechat-channels.md.
• 抖音: references/cases/douyin.md.
• 快手: references/cases/kuaishou.md.
• B站: references/cases/bilibili.md.
• 抖音电商/带货: references/cases/ecommerce-cases.md.

When a review uses a remote video URL, TikHub, OCR/ASR, Gemini, or another multimodal model, read references/video-evidence-integrity.md before acquiring media or assigning findings.

Local Evidence Search

Local search is available by default and does not require live-search permission because it reads only the static files shipped in this repository and makes no network request. Use tools/search_local_evidence.py when the user's symptom or question may span several reference files, or when a focused lookup is more efficient than reading an entire case catalog.

• Search only the published references/**/*.md and docs/sources.md corpus.
• Do not search local/; it is a developer-only, gitignored collection and is not present when users install the Skill.
• Cite every result by repository path, line, heading, and source type.
• Describe it as 本地静态证据, never as current/live platform evidence.
• No local match means only that the shipped corpus has no lexical match. It does not prove that the content is compliant or that no platform case exists.
• Local search may run automatically as part of an ordinary compliance review.

Optional Live Evidence

Live search is off by default. Do not search merely because a TikHub key, a platform browser plugin/Skill, or a general browser tool is installed. Run the normal review from official/static references and user-provided material unless the user explicitly asks to search current platform content, recent cases, or live discussion.

When the user explicitly requests live search, inspect the tools already available in the current environment and use a suitable available channel:

1. If the user names a provider or tool, use only that provider or tool. Do not silently substitute another source if it is unavailable.
2. Otherwise, prefer a target-platform-specific browser plugin or Skill when it can read the requested public content.
3. TikHub is an optional direct REST adapter. Use only the bundled tools/tikhub/bin/tikhub CLI or tools/tikhub/lib/tikhub_client.py, which call documented https://api.tikhub.io/api/v1/... endpoints. Never invoke a TikHub MCP server, mcp.tikhub.io, or a TikHub MCP tool. Use tools/xhs_dynamic_evidence.py diagnose for Xiaohongshu search, or cataloged Douyin REST endpoints for an explicitly requested video lookup.
4. A real browser automation tool such as agent-browser may be used when it can access the public search/results page. Respect login, access, CAPTCHA, rate-limit, and platform restrictions; do not bypass them.

The presence of a live-search channel is not authorization to call it. A live lookup may consume API quota, incur fees, or access a logged-in browser session, so it remains opt-in for each user request.

Before searching, report which channel will be used. If the requested channel is missing or fails, do not block the compliance review. Continue with official rules, static case files, and user-provided evidence, and state that live search was not performed or did not complete. Do not claim that no recent cases exist when the search itself failed.

Dynamic evidence must be reported under a separate 实时平台证据（可选） section with provider, search terms, target platform, sample time, content IDs or URLs, publish dates when available, and limitations. Ordinary creator posts and comments are discussion samples, not platform rules, and cannot determine the final severity by themselves.

Keep live API responses, downloaded media, signed URLs, model outputs, logs, and generated reports under a gitignored local evidence directory. Never commit API keys, cookies, authorization tokens, signed media URLs, raw creator data, or paid lookup results. Public commits may contain only reusable code, endpoint schemas, tests, Skill instructions, and static references.

Video File Analysis

When the user provides a local video file and asks whether it may violate a platform rule, use tools/analyze_video.py to prepare a local evidence package before assigning a risk level. This local media processing does not require live-search permission and must not trigger TikHub or browser search by itself. Follow references/video-evidence-integrity.md for remote acquisition, source ledger, OCR/ASR uncertainty, full-video model review, and mismatch handling.

The tool requires ffmpeg and ffprobe. It produces:

• manifest.json: media metadata, source hash, evidence coverage, exact frame timecodes, audio/subtitle status, technical signals, and text precheck hits.
• review_brief.md: a compact checklist and evidence index for the agent.
• frames/: opening 0-5s, scene-change, periodic, and ending frames.
• contact_sheets/: overview sheets for visual triage.
• audio.wav: a mono 16 kHz review copy when the video has an audio stream.
• Embedded/sidecar subtitle references and, only when explicitly enabled, a local Whisper transcript.

Example preparation command:

python3 tools/analyze_video.py --video <video.mp4> --platform <douyin> --output-dir <qa/video-evidence>

Use repeated --text-file <script-or-subtitle> arguments when a transcript, subtitle, or voiceover script is provided separately from the video.

Add --transcribe only when local Whisper is installed and transcription is needed. Whisper models may require a download; do not initiate a model download without telling the user. A supplied subtitle or script can be reviewed without Whisper, but subtitles do not prove that the audible voiceover is identical.

After preparation:

1. Read manifest.json and review_brief.md.
2. Inspect every contact sheet, then inspect individual original frames at all suspected timecodes. Pay special attention to the cover/first frame, first 0-5 seconds, scene changes, product/CTA segments, sensitive visuals, and end.
3. Review audible speech and BGM. Use a supplied transcript, an available audio review capability, or the optional local Whisper transcript. If audio cannot be reviewed, add it to 待核验 and do not return Pass.
4. Review visible subtitles, stickers, watermarks, QR codes, contact details, product claims, before/after imagery, AI labels, and privacy identifiers.
5. Map each observed element to the universal and target-platform references, with an exact timecode/frame pointer and a concrete remediation.
6. Treat automated black/silence detection and text claim matches only as precheck signals. The tool intentionally leaves the final verdict not_assigned; the agent must determine Pass/Low/Medium/High/Blocker from the evidence and clearly state coverage limitations.
7. If Gemini or another multimodal model is available and the user requests or authorizes it, submit the actual reviewable video file, preserve the raw structured response locally, and require an explicit successful result. Never treat a failure fallback, default Pass, contact-sheet-only analysis, or a model statement without confirmable timecode evidence as fact.

Frame sampling can miss very brief content. For high-risk, long, fast-cut, or regulated-domain videos, increase the sample count or inspect the full timeline. If only a remote URL is supplied, use an already available and authorized browser/download channel to obtain reviewable evidence. If the video cannot be accessed, state that it was not reviewed and request a local file, transcript, or screenshots; never infer a Pass from an inaccessible URL.

For new platforms, add one reference file under references/<platform>.md with:

• Platform scope and source date.
• Severity rules and hard blockers.
• Category checklist with platform article numbers or policy names.
• Common risky phrases, visuals, and behaviors.
• Safer rewrite/remediation patterns.

Do not overload this SKILL.md with platform rule catalogs; keep detailed platform material in references.

Review Workflow

1. Inventory the public surface

   • List every user-visible or user-audible element: picture, cover, first frame, title, subtitles, voiceover, BGM, captions, stickers, comments, private-message prompts, product card, external links, QR codes, account profile.
   • For videos, sample the first 0-5s, major scene changes, product/CTA sections, sensitive visuals, and ending.

2. Map content intent

   • Identify whether the content is education, entertainment, news/current affairs, product marketing, health/medical, finance, legal, relationship advice, minors, animals, violence, sexuality, or public-interest event coverage.
   • Flag regulated domains early: medical/health, finance/investment, legal services, fundraising, lottery/gambling, pet/vehicle trading, drugs/medical devices, health food, special medical formula food.

3. Run universal risk areas

   • Rights: copyright, low-effort搬运/二创, third-party watermarks, portrait/name/reputation/privacy, trademark/patent.
   • Sexual/lowbrow: nudity, body focus, sexual implication, sexual sounds/text, sex jokes, animal mating.
   • Violence/discomfort: gore, injury/death, surgery, abuse, bullying, horror, excrement/secretions, dense holes/insects, disturbing food/animals.
   • Illegal or harmful: gambling, pyramid schemes, controlled goods, illegal finance, fraud, fake cheating tools, dangerous stunts, minors' unsafe behavior.
   • Marketing: exaggerated claims, unverifiable data, absolute terms, fake authority, inconsistent price/gifts/link, nonofficial purchase channels, excessive product insertion.
   • Misinformation: outdated events as news, fake interviews, unknown-source emotional stories, celebrity rumors, AI/synthetic accident/war scenes without clear labeling, pseudoscience.
   • Inducement and diversion: forced likes/comments/follows/shares, curses/coercion, fake benefits, incomplete episodes, off-platform traffic, risky contact methods.
   • Public order and morals: discrimination, insults, public-order disruption, abnormal relationship sensationalism, family abuse, bad marriage customs.
   • Production quality: unreadable/misaligned subtitles, wrong aspect ratio, black screens, distorted visuals, audio dropouts, audio-video mismatch, invalid/unrelated links.
   • E-commerce product consistency (load references/douyin-ecommerce.md and references/qianchuan-low-quality.md):
     • Video-displayed product vs detail page: brand, style, color, pattern, shape consistent?
     • Voiceover spec vs SKU spec: quantity, weight, volume, size consistent?
     • Voiceover price vs linked product price: consistent?
     • Gift/red-packet/coupon conditions: complete and accurate?
     • Efficacy/performance claims: supported by qualification or evidence?
     • Source material authorization: video, BGM, images, fonts, portrait, brand assets.
     • 千川低质素材 patterns: product pile-up, 卖惨/演戏, picture carousel/big-character posters/high-saturation, black borders/watermarks/blur, unclear product subject, ingredient/size chart replacing product display.
     • Regulated category check: food, cosmetics, medical devices, drugs, maternal-child, pet, financial, education — qualification required.

4. Apply platform references

   • Cite platform category ids or article names where available.
   • When using recent examples, separate 官方/监管, 媒体转述, and 小红书讨论样本; do not treat creator comments as binding rules.
   • When optional live evidence is used, put it in a separate 实时平台证据（可选） section with search terms, sample date, note ids, and the evidence limitation.
   • Prefer the most specific matching category. If multiple categories apply, list all but mark the primary risk.
   • When the platform rule depends on account history or qualifications that are not available, mark 待核验.
   • Beyond official rules, platforms have many unwritten/隐形 rules. Treat creator-posted experience and comment-section discussion in the case files as a valuable supplement that surfaces these hidden enforcement patterns — but as symptoms and disputed edge cases, not as binding rules.

5. Classify severity

   • Blocker: clear illegal/high-risk violation, severe platform red line, likely takedown, user safety/property risk, unqualified regulated advice/marketing, porn/gambling/fraud, unmasked gore/death/sexual assault/minor harm, obvious unauthorized搬运, risky contact or off-platform diversion.
   • High: likely platform violation or strong enforcement risk; publish only after edits or proof is added.
   • Medium: ambiguous or context-dependent risk; revise, add context/disclosure, thicken masks, remove risky phrasing, or retain evidence.
   • Low: minor wording/UX/quality risk; monitor or polish.
   • Pass: no material risk found in the reviewed evidence.

6. Write concrete fixes

   • For audio risks: mute/beep exact ranges and rewrite subtitles/cards.
   • For visual risks: cut, replace, blur/mosaic, crop away, thicken masks, avoid using as cover/opening.
   • For claims: remove absolutes, add verifiable source/context, avoid guaranteed outcomes, disclose ad/marketing nature, align link price/gift/specs.
   • For regulated topics: remove advice/marketing, add verified qualification evidence, or convert to general non-prescriptive information.
   • For rights: replace with authorized/original material; attribution alone does not cure unauthorized use.
   • For inducement/diversion: remove coercive CTA, off-platform contacts, QR codes, risky private-message funnels, and fake benefits.

Evidence Standards

Every finding should include at least one evidence pointer:

• Video timecode or frame range.
• Script/subtitle line.
• Cover/title/caption/comment/product-link text.
• Screenshot/frame description.
• Missing proof: authorization, qualification, source, product price, activity scope, link consistency.
• For platform discussion evidence: keyword searched, note id or visible account, comment evidence if used, publish date, sample date, and whether it is official-account material or creator-side discussion.

Maintain a source ledger that separates user-supplied caption, platform/TikHub metadata, downloaded media, local frame/audio observations, OCR/ASR output, and multimodal-model output. If they disagree, report the mismatch and its possible technical causes as 待核验; do not decide which source is true or accuse a party of manipulation without independent evidence.

Do not report a violation solely because a topic is sensitive. Explain what visible/audible element creates the risk and which rule it maps to.

E-Commerce Entry Points

The same Skill supports three e-commerce review entry points. Choose the right one based on the user's stage:

入口 1: 发布前合规 (Pre-Publish Compliance)

Full review of the complete video package before publishing:

• Run all universal risk areas + e-commerce product consistency.
• Load all relevant platform references + e-commerce references.
• Output full Markdown report + optional machine-readable JSON.
• Blocker or unaccepted High = must not publish.

Use when: "交付前审核"、"发布前跑一遍"、"check this video before publishing".

入口 2: 选品前风险 (Pre-Selection Risk)

Quick risk screen before the creator commits to promoting a product:

• Check if the product category is regulated (medical, food, cosmetic, maternal-child, pet, financial, education).
• Flag missing qualifications.
• Flag high-risk claim patterns in the product detail page itself.
• Output a condensed risk score + blocker list.

Use when: "帮我看看这个品能不能带"、"选品前风险排查"、"这个商品有没有违规风险".

入口 3: 文案生成前风险 (Pre-Copywriting Risk)

Scan the product's claims before writing scripts:

• Extract every claim from the product detail page.
• Classify as safe/risky/forbidden.
• Generate a forbidden-expression list for the copywriter.
• Flag claims that need qualification proof before they can be used.

Use when: "写文案前帮我扫一遍商品"、"这个品的禁用词有哪些"、"生成前风险检查".

Machine-Readable Output

When the user asks for structured output, or when the report feeds into another Skill (e.g. capsule-cinema, commerce-selection-advisor, or a 飞书审批 flow), output both Markdown and JSON:

{
  "platform": "douyin",
  "content_type": "commerce_video",
  "risk_level": "high",
  "rule_version": "douyin-ecommerce-2026-08",
  "findings": [
    {
      "severity": "high",
      "rule_id": "商品三一致-规格",
      "evidence": "口播 00:12；详情页 SKU",
      "reason": "视频口播数量与商品详情页不一致",
      "fix": "统一口播、画面和详情页数量"
    }
  ],
  "pending_verification": ["商品资质", "素材授权"],
  "reviewed_at": "2026-08-11T10:00:00+08:00"
}

Field definitions:

• rule_id: use labels from the e-commerce reference files (e.g. 商品三一致-规格, 千川低质素材 卖惨/演戏/炒作, 电商功效声明 效果承诺).
• severity: blocker | high | medium | low | pass.
• evidence: video timecode, script line, detail page section, screenshot reference.
• pending_verification: list of items that could not be verified with the available evidence.

Report Format

Use this structure for serious reviews:

# 合规风险审核

- 平台: <platforms>
- 内容范围: <video/script/cover/copy/link/etc.>
- 结论: Pass | Low | Medium | High | Blocker
- 最高风险: <one sentence>

## 风险明细

| 等级 | 平台/条款 | 证据位置 | 风险说明 | 修改建议 |
| --- | --- | --- | --- | --- |
| High | 视频号 4.11.1 | 口播 00:12 / 标题 | 使用无法核验的销售数据 | 删除数字或补充可验证来源 |

## 待核验

- <authorization/qualification/link consistency/source provenance/etc.>

## 发布前复审清单

- [ ] 风险音频已消音或替换
- [ ] 风险字幕/封面/标题已改写
- [ ] 敏感画面已删除或充分打码
- [ ] 商品链接、价格、赠品、规格与视频一致
- [ ] 医疗/金融/法律等资质已核验或相关内容已移除
- [ ] 版权、肖像、隐私、商标授权已核验

For quick reviews, keep the same fields but collapse the table to bullets.

E-Commerce Extended Report

When reviewing带货 content, add a product-consistency table after the main risk table:

## 商品一致性审核

| 检查项 | 结果 | 证据 |
| --- | --- | --- |
| 视频展示商品 vs 商品详情页 | 一致 / 不一致 / 待核验 | 画面时间点；详情页截图 |
| 口播规格 vs SKU 规格 | 一致 / 不一致 / 待核验 | 口播时间点；SKU 页面 |
| 口播价格 vs 商品链接价格 | 一致 / 不一致 / 待核验 | 口播时间点；链接价格 |
| 赠品/红包条件是否完整 | 完整 / 缺条件 / 待核验 | 口播时间点；活动规则页 |
| 功效表达是否有资质证据 | 有证据 / 高风险 / 待核验 | 证据来源；资质文件 |
| 来源素材是否有授权 | 有 / 无 / 待核验 | 授权文件 |

## 千川低质素材检查

| 类别 | 是否触发 | 证据/说明 |
| --- | --- | --- |
| 商品堆砌式累加 | 是/否 | |
| 品牌/款式/颜色/形状不一致 | 是/否 | |
| 数量/重量/规格不一致 | 是/否 | |
| 卖惨/演戏/炒作/清仓话术 | 是/否 | |
| 价格/赠品信息不完整 | 是/否 | |
| 污垢/血腥/危险/恶俗/不适 | 是/否 | |
| 图片轮播/大字报/高饱和配色 | 是/否 | |
| 黑边/遮挡/水印/模糊变形 | 是/否 | |
| 商品主体不明确 | 是/否 | |
| 素材授权 | 是/否 | |

E-Commerce Tools

Two optional scripts supplement e-commerce reviews with automated checks. Use them when the user provides structured data (product JSON + claims JSON) or text to scan:

tools/compare_product_link.py

Paired comparison of product detail page info versus video claims:

python tools/compare_product_link.py \
  --product product_info.json \
  --claims video_claims.json \
  --format markdown|json

Output: brand, style/color, specs, price, gift, deadline consistency results + regulated-category risk flag + 千川 checklist.

tools/extract_claims.py

Scan script/subtitle/copy text for risky claims before writing or reviewing:

python tools/extract_claims.py \
  --file script.txt \
  --format markdown|json

Output: forbidden expressions (must delete), risky expressions (should rewrite), and items needing evidence — each with a suggested fix.

Both tools output JSON for machine-to-machine handoffs (e.g. capsule-cinema, commerce-selection-advisor, 飞书审批).

Video Workflow Integration

When reviewing a final video package:

• Prepare local video evidence with tools/analyze_video.py when a video file is available.
• Keep generated frames, audio, transcripts, and manifests under qa/ or internal/; they are review artifacts, not publishable assets.
• Save the report as qa/compliance_review.md when the project has a release package.
• If the package separates public/internal files, keep risk notes in qa/ or internal/, not in publishable public/ copy.
• Update the release manifest or handoff notes with the compliance report path and final risk level.
• Do not call a video final if there is any unresolved Blocker or unaccepted High risk.

Common Mistakes

• Treating subtitle rewrite as enough while risky audio remains audible.
• Reviewing only the script and missing cover/opening-frame risks.
• Assuming public material is safe because it came from another platform.
• Treating 小红书 creator notes or comments as official platform rules instead of discussion samples.
• Using "仅供参考" to keep medical, financial, legal, or guaranteed-effect claims that still require qualification or proof.
• Leaving product prices, gifts, quantities, or activity deadlines inconsistent with the linked item.
• Hiding uncertain source provenance instead of marking it 待核验.
• Reviewing带货 content without loading e-commerce references (douyin-ecommerce.md, qianchuan-low-quality.md, ecommerce-claims.md).
• Treating product claims from the detail page as verified facts without checking qualifications.
• Missing 千川低质素材 patterns (picture carousel, big-character posters, black borders) that don't look like rule violations but trigger promotion rejection.
• Skipping the product-consistency check: video display, voiceover, subtitles, and product link must all agree on brand, specs, price, and gifts.
• Marking regulated-category products (食品, 化妆品, 医疗器械, 母婴, 宠物, 金融) as Pass without qualification evidence.
• Calling sampled keyframes "逐帧语义审核" or claiming every decoded frame was reviewed when only contact sheets or periodic frames were inspected.
• Treating OCR, Whisper, or model-inferred speech as exact dialogue without confirming visible subtitles, audio, or a supplied transcript.
• Treating a preallocated, partial, stalled, or ffprobe-invalid download as a complete original video.