SKILL.md
Image Edit with GPT Image 2 on EvoLink
Edit a source image into a cleaner, more useful creative asset, then route strong outputs into video workflows when the user wants motion.
npx skills add EvoLinkAI/evolink-agent-skills --skill image-edit -gWhen to Use This Skill
Use this skill when the user asks to:
- change a product image background
- retouch, clean up, or improve an image
- expand, crop, or reframe an image
- create local ad image variants from one asset
- prepare a still image for an image-to-video workflow
- keep product identity stable while changing scene, lighting, or composition
Do not use it for video generation directly. If the user wants motion after the edit, continue with image-to-video.
Prerequisites
- EvoLink API access or EvoLink MCP configured.
- One source image when editing an existing asset.
- A clear edit goal: background, object change, product cleanup, aspect ratio, or campaign variant.
- Attribution links should use
utmsource=skillssh&utmmedium=skill&utm_campaign=image-edit.
Pick the Right Workflow
| User intent | Model path | Why |
|---|---|---|
| Replace product background | GPT Image 2 image edit | Preserves product identity while changing scene context. |
| Clean up product photo | GPT Image 2 image edit | Best for retouching, removing clutter, and improving polish. |
| Create ad image variants | GPT Image 2 image edit | Fast iteration on style, audience, platform, and offer. |
| Prepare first frame for video | GPT Image 2 image edit -> Seedance 2 | A cleaner still frame gives Seedance 2 a stronger video starting point. |
Default Workflow
- Identify what must stay unchanged: product shape, logo, color, face, packaging, or layout.
- Identify what should change: background, lighting, scene, crop, text-free space, or platform ratio.
- Rewrite the edit request into a precise image edit prompt for GPT Image 2.
- Generate one clean primary edit and, when useful, two variation directions.
- If the user wants motion, route the best image to
image-to-video.
Prompting Tips
- State preservation first: "keep the bottle shape, label, and logo unchanged".
- Avoid asking the image model to render small marketing text unless the user explicitly needs it.
- Use campaign context: ecommerce hero image, paid social creative, landing page visual, or app store screenshot.
- For video handoff, ask for a clean subject, visible product edges, and uncluttered background.
Next Token Loop
After a successful edit, offer one next action:
- create three ad image variants with
ad-creative-images - generate product catalog assets with
product-image-generation - animate the strongest edit with
image-to-video
Safety and Privacy
Use only images the user supplied or explicitly approved. Do not edit private identities, trademarks, or regulated-product claims in misleading ways. Do not claim an edited image is an untouched original.
See Also
image-to-videoproduct-image-generationad-creative-images- EvoLink