Skill: fashion-video-creator

Overview

Generate complete assets for fashion e-commerce videos from scratch: a virtual model image (Seedream text-to-image) + a Seedance 2.0 video prompt with dialogue, actions, and scene + an operator SOP for the Jimeng platform. Supports single and batch modes.

When to Activate

User query contains any of:

• - "穿搭视频", "outfit video", "生成模特", "虚拟模特"
• "seedance prompt", "穿搭创作", "带货视频", "带货短视频"
• "帮我做个穿搭视频", "批量生成穿搭", "batch prompt"
• "帮我拍个带货视频", "我有件衣服想做个展示视频", "做个穿搭短视频"
• "generate fashion video", "create outfit content"

Do NOT activate for:

• - Analyzing / reverse-engineering existing videos -> use INLINECODE0
• "这个视频怎么拍的", "帮我复刻这个视频" -> use INLINECODE1
• Pure image generation without video prompt -> direct Seedream API call
• Non-fashion video generation -> not applicable

Prerequisites

This skill requires a Volcano Engine (火山方舟) account with Seedream model access.

CODEBLOCK0

Clarification Flow

Phase 1: API Key Acquisition

Ask these questions IN ORDER. Use plain language — the user may not be technical.

Q1: Image Generation Service

"你需要先有一个能生成图片的 AI 服务账号。目前这个工具使用的是火山方舟平台的 Seedream 模型。

你有火山方舟的账号吗？如果有，请提供你的 API Key（在火山方舟控制台 -> API Key 管理中可以找到）。"

If user has no account -> STOP. Guide them to register at 火山方舟. Do NOT proceed without API key.

Q2: Model Version

"火山方舟上有两个版本的图片生成模型：

• - Seedream 5.0（推荐）：理解能力更强，生成的人像更自然，每张约 0.22 元
• Seedream 4.5：可以更精细地控制细节，每张约 0.32 元

你开通了哪个？需要提供模型ID或接入点ID。"

Phase 2: Mandatory Recommendations

MUST show before proceeding. These are not optional tips — violating them degrades output quality:

CODEBLOCK1

Phase 3: Mode Selection

"你要生成几条视频？单条还是批量？"

Phase 4: Creative Parameter Collection

Collection strategy — ask in 2 rounds, never more than 4 questions per round:

Round 1 (required — must have before executing):

1. 1. Model gender (female/male) — default: female
2. Garment type (dress/top/pants/jacket/suit/casual) — NO default, must ask. WHY: "default" type produces generic dialogue without product-specific selling points. Defensive fallback: If user declines to specify after 2 attempts, fall back to "default" with warning: "Using generic dialogue — conversion effectiveness will be lower than type-specific scripts."
3. Scene (11 presets or custom text) — default: modern_apartment

Round 2 (enhanced — use defaults if not stated):

1. 4. Model appearance: preset (9 templates) / body params / custom desc — default: asianfemaleslim
2. Realism (0-100) — default: 40
3. Style: realism / pixar / anime_realistic — default: realism
4. Dialogue style: natural / professional / enthusiastic — default: natural
5. Camera: vlog / pro / static — default: vlog
6. Target audience: auto / female / male — default: auto
7. Duration (5-40s) — default: 10
8. Language: zh / en — default: zh
9. Task label (batch only)

Batch shortcut: If all tasks share config, ask once and apply to all. Only collect per-task differences.

Batch-Specific Recommendations

When batch mode selected, ADDITIONALLY show:

CODEBLOCK2

Core Workflow

Step 0: Verify API Key (mandatory, never skip)

Platform-adaptive verification:

If bash/Python available (Claude Code, terminal):
CODEBLOCK3

If no code execution (Claude.ai web, chat-only):
Trust the user-provided key and proceed. Mark internally: api_verified: false.
If Step 3 (image generation) fails with 401 -> surface the error then.

• - Key valid -> proceed to Step 1
• Key invalid (401/403) -> report error, ask user to verify. Do NOT proceed.
• No code execution -> proceed with unverified key, validate on first API call.

Single Mode

CODEBLOCK4

Batch Mode

CODEBLOCK5

Output Validation (mandatory, never skip)

Before delivering results, verify ALL:

• - [ ] Model image is 9:16 portrait with full body visible (head to feet)?
• [ ] Prompt contains @image1 (garment) and @image2 (model ref)?
• [ ] Dialogue matches the selected garment type (not generic "default")?
• [ ] SOP specifies correct image upload order (image1=garment FIRST)?
• [ ] If chained: every segment has continuation prompt + dialogue preview?
• [ ] Brand constraint present: "Do not alter clothing pattern, color, texture or style"?

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

Error Handling

Failure	Detection	Action
No API key	ARKAPIKEY empty or missing	STOP. Guide user to 火山方舟 console. Do NOT proceed.
Invalid API key

401/403 from API | Report error. Ask user to verify key. Do NOT retry with guessed keys. | | Seedream timeout | No response in 300s | Retry once with same prompt. Still fails -> report with the prompt used so user can try manually. | | No image in response | API returns empty data | Simplify prompt (remove extras), retry. Still empty -> report error. | | Batch task fails | Exception during prompt assembly | Mark task as failed with error message. Continue remaining tasks. Report partial results at end. | | Invalid garment type | User provides unknown type | Map to closest valid type or ask for clarification. Valid: dress/top/pants/jacket/suit/casual/default. |

Degraded Mode: Prompt-Only (when image generation fails)

If Seedream API is unreachable or fails after retry, do NOT block the entire workflow.
Skip image generation, still deliver prompt + SOP. Mark output clearly:

CODEBLOCK6

This ensures the user gets the valuable prompt + SOP even without the AI-generated model image.

Usage Example

Input: "帮我做一个女性穿搭视频，裙子类型，现代公寓场景"

Resolved params: gender=female, garment=dress, scene=modern_apartment, realism=40, style=realism, dialogue=natural, camera=vlog, audience=auto->female, duration=10, lang=zh

Output 1 — Model Image: 720x1280 PNG (asianfemaleslim preset, realism=40)

Output 2 — Seedance Prompt:
CODEBLOCK7

Output 3 — Operator SOP: Step-by-step for 即梦 platform (upload order, settings, quality checklist)

Domain Knowledge Role Declaration

The reference files contain dialogue scripts, model presets, API specs, and prompt templates.
Their role is to assist prompt assembly — providing the raw materials that get composed into Seedance prompts.
They do NOT replace the execution workflow. Never output reference content directly as the final answer.
Always assemble through the workflow: collect params -> build prompt -> validate -> deliver.

References

File	Purpose	When to read
references/model-presets.md	9 model presets, body params, 17-level realism anchors, scene/camera/style presets	Step 2: building model prompt
references/seedream-api.md

Seedream 4.5/5.0 API endpoint, request format, cost, crop logic | Step 3: calling API | | references/prompt-assembly.md | Prompt composition order, multi-segment chaining constants, SOP templates | Step 4-6: assembling prompt and SOP | | references/dialogue-library.md | Complete dialogue scripts: 7 garment types x 2 genders x 2 languages (28 scripts total), openings, closings, actions | Step 4: composing dialogue content |