Skill: viral-video-replicator

Overview

Reverse-engineer reference videos (e.g., competitor viral content) into replicable Seedance 2.0 prompts. The pipeline: FFmpeg frame extraction -> contact sheet grids -> audio extraction + ASR transcription -> Vision LLM structured analysis -> Seedance prompt assembly with optional material replacement (face/body/clothing). Supports single and batch modes.

When to Activate

User query contains any of:

• - "视频复刻", "视频逆向", "反编译视频", "复刻爆款"
• "分析这个视频", "replicate this video", "video analysis"
• "批量分析", "批量复刻"
• "这个视频怎么拍的", "帮我分析一下这个爆款", "我想拍一个类似的视频"
• "reverse engineer this video", "analyze this fashion video"

Do NOT activate for:

• - Creating fashion videos from scratch (no reference video) -> use INLINECODE0
• "帮我做个穿搭视频", "生成模特图" -> use INLINECODE1
• Pure video editing / trimming -> not applicable
• Non-fashion video analysis -> not applicable

Prerequisites

Local tools (REQUIRED):
CODEBLOCK0

Cloud APIs (collected via clarification):
CODEBLOCK1

Clarification Flow

Phase 1: API Key Acquisition

Ask IN ORDER. Use plain language — explain WHY each service is needed.

Q1: Vision Analysis (REQUIRED)

"分析视频内容需要一个能'看懂图片'的 AI 模型。

它会看视频的截图，识别出人物长相、服装细节、场景布局、动作时间轴。

你有火山方舟的账号和 API Key 吗？还需要视觉模型的ID。"

If no API key -> STOP. Guide user to 火山方舟. Do NOT proceed.

Q2: Speech Transcription (CONDITIONAL)

"参考视频里的人有说话吗？

如果有对话，需要用语音转文字来提取台词 — 这样复刻出的视频才能有完整的对白内容。

没有对话的纯画面视频可以跳过这步。"

If yes -> ask for ASRACCESSTOKEN.

Q3: Audio Storage (CONDITIONAL — only if Q2 = yes)

"语音转文字需要通过云存储传输音频文件。

需要火山引擎对象存储(TOS)的 4 个信息：Access Key, Secret Key, Bucket, Region。"

If user has ASR but no TOS -> warn: "没有 TOS 则 ASR 无法工作，等同于没有语音转录。"

Phase 2: Mandatory Recommendations

MUST show. Each item has WHY explanation:

CODEBLOCK2

Phase 3: Mode Selection

"你要分析几个视频？单个还是批量？"

Q5: Replicate Mode (per video if batch)

"你想怎么复刻？

• - 精确复刻: 逐帧分析每个细节，尽可能1:1还原
• 提取改写: 提取爆款节奏和逻辑，用新方式重新演绎"

Q6: Material Replacement (per video if batch)

"要替换视频中的哪些元素？

• - 不换（纯复刻）
• 换人脸/身材（上传模特参考图）
• 换衣服（上传商品图）
• 都换（上传模特图 + 商品图）"

Batch-Specific Recommendations

CODEBLOCK3

Four Replacement Modes

Mode	What User Uploads	@image Tags in Prompt	What Gets Replaced
clone	Nothing	None (pure text)	Nothing — exact replication
face_swap

Face/body reference | @image1 = face ref | Person replaced, clothing preserved | | outfit_swap | Garment product image | @image1 = garment | Clothing replaced, person preserved | | full_swap | Garment + face reference | @image1 = garment, @image2 = face ref | Both replaced |

Mode auto-determination:
CODEBLOCK4

Core Workflow

Step 0: Environment Check (mandatory, never skip)

CODEBLOCK5

• - Returns version -> proceed to Step 1
• INLINECODE2 -> guide install (brew/apt/choco). Still fails after install ->

Soft fallback: Ask user: "FFmpeg 不可用，你能手动提供视频截图和音频文件吗？" If user provides frames manually -> skip FFmpeg steps, proceed from Step 4 (Vision analysis) with user-provided images. Quality warning (MUST show to user): "手动截图模式下分析质量会显著降低：无精确时间戳标注、无均匀3fps采样、帧数可能不足导致动作时间轴不准确。建议安装 FFmpeg 以获得最佳效果。" If user cannot provide frames -> STOP. FFmpeg is required for automated extraction.

Step 0b: Verify API Key (before reaching Step 4)

Validate ARKAPIKEY early to avoid wasting FFmpeg processing time on an invalid key:

If bash/Python available:

resp = httpx.get(f"{ARK_API_BASE}/api/v3/models",
                 headers={"Authorization": f"Bearer {ARK_API_KEY}"}, timeout=10)

• - 200 -> proceed
• 401/403 -> STOP. Key invalid. Fix before continuing.

If no code execution: Trust user-provided key, validate on first Vision API call.

Single Mode

CODEBLOCK7

Batch Mode

CODEBLOCK8

Output Validation (mandatory, never skip)

Before delivering results, verify ALL:

• - [ ] Analysis JSON is valid and contains all required fields?
• [ ] Prompt correctly uses @image tags matching the replacement mode?
• [ ] If clone mode: prompt has NO @image references (pure text)?
• [ ] If outfitswap/fullswap: prompt includes "Do not alter clothing pattern, color, texture or style"?
• [ ] If has_speech: dialogue content is present in prompt (not empty)?
• [ ] SOP upload instructions match the number of images for this mode?
• [ ] Replacement summary correctly lists what was preserved vs replaced?

Any NO -> fix before delivering. Do NOT send unvalidated output.

Error Handling

Failure	Detection	Action
FFmpeg not installed	INLINECODE3	STOP. Provide install command. Do NOT proceed.
No API key

ARKAPIKEY empty | STOP. Guide user to 火山方舟. Do NOT proceed. | | Vision model error | 4xx/5xx from API | Report error with model ID used. Suggest checking model availability. | | Vision returns invalid JSON | JSON parse fails | Retry once with same grids. Still fails -> report raw response for debugging. | | Frame extraction fails | FFmpeg non-zero exit | Check video format. Try re-encoding. Report if still fails. | | No audio track | extract_audio returns None | Skip ASR. Proceed with visual-only analysis. Note in output: "No audio detected." | | TOS upload fails | Upload exception after 2 retries | Skip ASR. Proceed visual-only. Warn: "Audio transcription unavailable — dialogue will be missing." | | ASR timeout | No result after 120s | Skip transcript. Proceed visual-only. Warn: "Speech transcription timed out." | | ASR silent audio | Status 20000003 | Normal — video has no speech. Proceed with visual-only. | | Video too large | >200MB | Reject immediately. Ask user to compress or trim. | | Batch video fails | Exception during pipeline | Mark failed with error. Continue remaining. Report partial results. |

Degraded Modes (graceful degradation chain)

Failure Point	Degraded Mode	What User Still Gets	Quality Impact
ASR fails (TOS/timeout)	Visual-only analysis	Prompt with visual descriptions, no dialogue	~50% fidelity — all spoken content lost
Vision exact mode fails

Auto-retry with rewrite mode | Flat analysis (less precise) | ~70% fidelity — loses nested structure (clothing/scene subfields) | | Vision rewrite also fails | Return raw materials | Frame grids + transcript for manual analysis | ~20% — no automated analysis, user must write prompt manually | | Seedance prompt assembly fails | Return analysis only | Analysis JSON + transcript | ~30% — user has data but no ready-to-use prompt | | FFmpeg unavailable (user provides screenshots) | Manual frame mode | Analysis from user-provided images | ~40% — no timestamps, uneven sampling, incomplete frame coverage |

Always prefer delivering partial results over delivering nothing. Every degraded output MUST clearly state: (1) what is missing, (2) why, and (3) the estimated quality impact.

See references/fallbacks.md for detailed recovery procedures per failure case.

Usage Example

Input: "帮我复刻这个爆款视频，换成我的衣服" + uploaded video (15s, 720p) + uploaded garment image

Resolved: mode=exact, replacement=outfitswap (garmentref provided, no face_ref)

Output 1 — Structured Analysis:
CODEBLOCK9

Output 2 — Seedance Prompt (outfit_swap):
CODEBLOCK10

Output 3 — Transcript: "姐妹们你们快看...这个面料是醋酸缎面的..."

Output 4 — SOP: outfit_swap mode, 1 image upload (@图片1=garment)

Output 5 — Replacement Summary: garmentreplaced=true, originalpreserved=[face, body, scene, actions, dialogue, camera]

Domain Knowledge Role Declaration

The reference files contain FFmpeg specs, ASR protocols, Vision prompts, and prompt assembly templates.
Their role is to assist pipeline execution — providing exact API formats, analysis schemas, and assembly rules.
They do NOT replace the execution workflow. Never output reference content directly as the final answer.
Always execute: extract frames -> transcribe -> analyze -> assemble -> validate -> deliver.

References

File	Purpose	When to read
references/frame-extraction.md	FFmpeg filter chain, grid stitching, audio extraction specs	Step 2: extracting frames and audio
references/asr-pipeline.md

TOS upload protocol, Seed-ASR-2.0 submit/poll API | Step 3: transcribing audio | | references/vision-analysis.md | Vision LLM prompts for exact and rewrite modes, output schemas | Step 4: analyzing video | | references/reverse-prompt.md | 4-mode prompt assembly, clothing generalization map, SOP templates | Step 6-7: building prompt and SOP | | references/fallbacks.md | 8 failure cases with recovery procedures and degradation chain | On any error during Steps 2-8 |