| name | ecom-shoe-image-shotlist-github |
| description | Footwear e-commerce photography expert. Given a white-background shoe image, automatically analyzes product (compress_product_img.py → Read tool → LLM vision) → selects scene from professional scene library → generates 6 advertising-quality hero images in 2-batch parallel pipeline (batch 1 baseline model → batch 2 remaining 5 images) using three-step pattern (submit → poll → download/upload to CDN) covering on-feet full-body shots, on-feet close-ups, and lifestyle still-life shots → optional video generation for 5 images with highlight reel. Use when the user mentions "shoe photography", "e-commerce hero image", "product shot", or uploads a shoe product image. |
Footwear E-Commerce Hero Image + Video One-Stop AIGC Generation
Input a white-background shoe image to automatically generate a set of 6 advertising-quality hero images (on-feet full-body shots, close-up shots, lifestyle still-life shots), and optionally generate short videos compiled into a highlight reel with background music. Supports athletic shoes, fashion shoes, and children's shoes; Asian and Western models.
Load sub-files on demand (do not load all at once):
- When selecting scenes → read
SCENE-LIBRARY.md
- When referencing complete prompt formats → read
PROMPT-EXAMPLES.md
Overall Style Tone
- Quality positioning: E-commerce fashion advertising campaign level, referencing Nike / Adidas / Zara / Massimo Dutti
- Lighting: Natural light preferred (morning light / evening soft light / indoor window light), avoid flat lighting, pursue three-dimensionality
- Color tone: Fashion shoes → premium minimalist cool/warm tone; Athletic shoes → energetic tone; Children's shoes → bright and fresh tone; natural and moderate color, no heavy filters
- Texture: Subtle film grain, close to fine film photography feel
Input Collection (Must Confirm Before Starting)
Required:
- White-background product image path
- Shoe type: athletic shoes / fashion shoes / children's shoes
- Model type: Asian/Western + male adult / female adult / boy / girl
Optional:
- Project name (if not provided, auto-use timestamp
YYYYMMDD_HHMMSS)
- Scene requirements (user-specified takes priority; otherwise read
SCENE-LIBRARY.md for auto-selection)
⚠️ When shoe type and model type are already provided in the resume prompt (via static intro follow-up events), parse the values directly and proceed — do NOT ask these questions again. Only use AskUserQuestion when inputs are genuinely missing (e.g., user uploaded a photo mid-conversation without going through the intro flow). See CLAUDE.md "Input Collection Rules" for the exact template.
🚨 After calling AskUserQuestion for shoe type / model type, you MUST end your turn and wait for the user's actual selection. NEVER auto-select or proceed to product analysis before the user responds.
Standard Image Set (Fixed 6 Images)
| # | Type | Key Specs |
|---|
| 01-02 | On-feet full-body shots | With face, medium/3-4 composition, 35-50mm, f/2.8-4, face must be visible |
| 03-04 | On-feet close-up shots | Below-knee close-up, 85-105mm, f/1.8-2.8, strong background blur; 03/04 must clearly differ in angle/foot pose |
| 05-06 | Lifestyle still-life | Product only, no people; same scene style, 05/06 must clearly differ in position/angle/perspective/shoe placement, no switching to completely different scene type |
Global Consistency Constraints:
- 01-04: Same model appearance, same outfit (including sock color and style), same scene tone
- 03-04 blurred background color tone must match 01-02 scene
- Lifestyle shots (05-06) MUST use the exact same scene location as on-feet shots (01-04) — same background, same surface material, same lighting environment; this is a hard constraint, not optional
- Image 02: Only reference the model appearance from image 01, must simultaneously satisfy:
- Pose/action must be clearly different (01 standing → 02 walking/turning; 01 frontal → 02 side/back; never reuse similar poses)
- Scene makes minor variations within the same style (no switching to completely different scene type; e.g., if 01 is outdoor street, 02 cannot be indoor studio; differences should be in: angle/position/lighting direction/background local elements, at least 3 must differ)
- Lifestyle shot hard constraint: No people, hands, feet, or pets in frame
- Full-body shot hard constraint: No other people in background
Object Scale (must declare in lifestyle shot prompts):
- Adult shoes approx. 25-30cm, children's shoes approx. 15-20cm
- Shoe-to-surrounding-object size ratio must be realistic
Tool Paths
All scripts are located in the production base directory:
/home/bun/.claude/apps/app-ecom_shoe_image_shotlist/.claude/skills/ecom_shoe_image_shotlist/scripts/
Product analysis:
Use your visual capabilities (ViewImage) to perform structured analysis on the provided image.
Image generation:
Use GenerateImage tool to concurrently generate all 6 images.
Video prompt generation:
Generate video prompts based on the templates in PROMPT-EXAMPLES.md.
Video generation:
Use GenerateVideo tool to concurrently generate all video clips.
Highlight reel merge: merge_highlight_video.py
Merges 5 video clips with xfade transitions + background music → uploads final video to CDN.
BGM auto-selected from music library based on shoe type.
Background music library:
/home/bun/.claude/apps/app-ecom_shoe_image_shotlist/.claude/skills/ecom_shoe_image_shotlist/assets/background-music.json
Safety: No hardcoded API Keys; all models are virtual/synthetic.
Image Generation Workflow
2-Batch Parallel Strategy: Image 01 must be generated first (baseline model reference), then images 02-06 are generated in parallel as a single batch.
-
Create project folder: mkdir -p projects/ecom-shoe/[project name]
-
Confirm reference image + required info
-
Select scene (user-specified takes priority; otherwise read SCENE-LIBRARY.md for smart selection, avoid reusing recent scenes)
-
Write all 6 prompts, then generate in 2 batches:
Batch 1 — Image 01 (baseline model, must complete first):
- Image 01: White-background product as reference, generate normally (serves as model appearance reference baseline for all subsequent images)
Batch 2 — Images 02-06 (5 tasks in parallel, after Batch 1 completes):
- Image 02:
img_urls: [WHITE_BG_URL, IMAGE_01_CDN_URL] dual reference, only reference model appearance from image 01; before writing prompt, clarify: ① different action/pose from image 01 (e.g., 01 standing → 02 walking; 01 frontal → 02 side), ② different specific position/angle/lighting within the same style, never switch to completely different scene type
- Images 03-04:
img_urls: [WHITE_BG_URL, IMAGE_01_CDN_URL] dual reference, ensure outfit/socks match image 01, strong background blur; 03/04 shooting angles must clearly differ (e.g., 03 side → 04 diagonal front; 03 static → 04 tiptoe/stepping), never reuse similar poses
- Images 05-06:
img_urls: [WHITE_BG_URL] single reference (no model), lifestyle shots; the scene MUST be the exact same location as images 01-04 (same background, same surface/floor material, same props, same lighting direction) — only the shoe placement, angle, and composition differ between 05 and 06. Before writing prompt: ① explicitly state the same scene elements from 01-04 (e.g., "same wooden café steps, same grey wall, same window light from left"), ② 06 must clearly differ from 05 in placement/angle (e.g., 05 top-down flat → 06 eye-level 45° angled; 05 single front view → 06 pair combination). Never invent a new or different scene type.
- If referencing complete prompt format → read
PROMPT-EXAMPLES.md
Sequential per-image generation is strictly forbidden — always use batch submission.
-
After generation, check file size; if over 3MB, compress (adjust JPG quality)
Prompts must include: Model description + shoe description + scene setting + action/pose + composition/angle + lens language + lighting texture + color tone + style reference
Quality Checklist (Review Each Image)
Advertising campaign quality (highest priority):
- ✓ Matches Nike/Zara/Massimo Dutti campaign quality, subtle film grain
- ✓ Rich light-shadow layering, no flat lighting; natural color tone, no heavy filters
- ✓ Natural model movement, professional poses, appropriate expressions (avoid stiff/wooden)
Basic items:
- ✓ Model appearance/outfit (including socks)/scene fully consistent across images 01-04
- ✓ Model face clearly visible in full-body shots
- ✓ Lifestyle shots have realistic object proportions, natural contact shadows, no floating effect
- ✓ Images 05-06 scene is identical to images 01-04: same background, same surface material, same lighting direction — if they look like a different location, reject and regenerate
- ✓ Image 06 and 05 have clearly different shoe placement/angle/composition (no reuse of similar framing)
- ✓ Image 02 and 01 pose/action clearly different (cannot reuse similar poses)
- ✓ Image 04 and 03 shooting angle/foot pose clearly different (cannot reuse similar angles)
Video quality review (manual confirmation required after generation):
- ✓ Shoes visible throughout the video, not out of frame
- ✓ No upward push causing shoes to disappear
- ✓ Full-body shots (01/02 videos) show from head to toe throughout
- ✓ Close-up shots (03/04 videos) lock onto lower leg + shoe area
- If failed → adjust prompt with stronger constraint words and regenerate that video
- ✓ Image 02 and 01 have same scene style, only minor differences in position/angle/lighting/background elements (no scene type switching)
- ✓ 1500×2000px, < 3MB, JPG
User Quality Review & Feedback Loop
After all 6 images are generated, proactively ask:
"Please review these 6 images. Are they up to standard? If any are not, please indicate the number and reason."
- All pass → proceed to video generation workflow
- Some fail → only regenerate failing images (keep passing ones), adjust prompts targeting the issues, ask again
- Same image fails 2 consecutive times → analyze root cause, adjust more elements like scene/angle/lighting
Video Generation Workflow (Optional After Images Pass)
Ask user if they want to generate videos. Only generate videos for images 01-05, skip image 06.
Step 1: Generate Video Prompts
Generate video prompts based on the templates in PROMPT-EXAMPLES.md.
Core principles:
- Models must have natural movement (turning/weight shift/arm swing/leg micro-movement), cannot be static
- Shoes must not deform, can rotate slightly and slowly (≤30 degrees, slow and steady)
- Shoes must be visible throughout the video, this is a hard constraint, non-negotiable
- Camera movement strict restrictions:
- Prohibited: pushing upward past waist (causes shoes to leave frame)
- Prohibited: large push-pull (over 10% of frame range)
- Allowed: slight pan (left-right ≤5%), slight downward push (toward shoe area), slow pull-back
- Full-body shots (01/02): camera stays at medium shot, ensuring head to toe visible throughout
- Close-up shots (03/04): camera locks on lower leg + shoe area, no upward drift
- Prompts must explicitly include:
camera stays on shoes throughout, shoes always visible in frame, no upward pan
After showing 5 prompts, review each one: does each prompt include shoes-always-visible constraint words (shoes always visible, no upward pan), then ask user to confirm before starting generation.
Step 2: Create Prompt Text Files
Create inside project folder (same filename as image, with .txt extension):
01-onfeet-full-01.txt / 02-onfeet-full-02.txt / 03-onfeet-closeup-01.txt / 04-onfeet-closeup-02.txt / 05-lifestyle-scene-01.txt
Step 3: Generate Videos One by One
Use GenerateVideo tool to concurrently generate all video clips:
prompt: corresponding video promptfirst_img_url: the URL returned when that image was generated (if only local file available, first upload using file_upload_to_obs_sync to get URL)- Fixed parameters:
duration=5, ratio="3:4", resolution="1080p"
Download each video immediately after completion, output: 01-onfeet-full-01.mp4 … 05-lifestyle-scene-01.mp4
Step 4: Ask if Merging Highlight Video (use AskUserQuestion)
Use AskUserQuestion to ask whether to merge — do NOT use plain text. One question, two options:
- Option 1: "Yes, merge with BGM" — merge 5 clips into a ~15s highlight reel with background music
- Option 2: "No thanks" — skip merge
If the user selects yes, call the merge_highlight_video.py script:
cd /home/bun/.claude/apps/app-ecom_shoe_image_shotlist/.claude/skills/ecom_shoe_image_shotlist/scripts && python3 merge_highlight_video.py \
--videos "VIDEO_01_PATH" "VIDEO_02_PATH" "VIDEO_03_PATH" "VIDEO_04_PATH" "VIDEO_05_PATH" \
--shoe-type SHOE_TYPE \
--output-dir "PROJECT_OUTPUT_DIR"
- Replace
VIDEO_XX_PATH with the local file paths from Video Step 3 (e.g., projects/ecom-shoe/outputs/01_video.mp4)
- Replace
SHOE_TYPE with one of: athletic, fashion, children
- Replace
PROJECT_OUTPUT_DIR with the project output directory
- The script automatically:
- Trims each clip to 3.5 seconds
- Applies fade crossfade transitions (0.5s) between clips
- Selects background music from the library matching the shoe type
- Mixes BGM with original audio (BGM fades out at the end)
- Uploads the final highlight reel to CDN
- On success: prints
[MERGE_RESULT] followed by a JSON object with r2_url (CDN URL), duration, bgm_label, file_size_mb
- Final duration ≈ 5×3.5 − 4×0.5 = 15.5 seconds
- Optional parameters:
--clip-duration 3.5, --xfade-duration 0.5, --bgm <path_or_url>
Naming Convention
Project folder: projects/ecom-shoe/[project name or timestamp]/
Images:
01-onfeet-full-01.jpg 02-onfeet-full-02.jpg
03-onfeet-closeup-01.jpg 04-onfeet-closeup-02.jpg
05-lifestyle-scene-01.jpg 06-lifestyle-scene-02.jpg
Videos (inside project folder):
01-onfeet-full-01.mp4 … 05-lifestyle-scene-01.mp4
01-onfeet-full-01.txt … 05-lifestyle-scene-01.txt (prompt files)
background_music.mov
final_video_with_music.mp4
