image-review

Image Review

When to Use

Use this skill when the user provides an image or image path and asks whether it is good enough, matches the prompt, or needs regeneration/editing.

Use especially when the image includes:

• text, logo-like elements, product details, people, hands, faces, or complex layouts.
• commercial, brand, social, presentation, ad, ecommerce, or print use.
• reference-image consistency requirements.
• safety, copyright, privacy, or misleading-content risks.

Do not invent an image review if the image cannot be viewed or read. Ask for an upload/path or state that only a generic checklist can be provided.

Workflow

1. Gather context:

• original user request, prompt, target use, platform/format, reference images, and output image.

1. Inspect what is visible:

• subject, composition, text, lighting, colors, artifacts, crop, resolution, and file constraints.

1. Score the image:

• Prompt adherence.
• Technical quality.
• Composition and aesthetics.
• Fitness for target use.
• Text and detail reliability.
• Reference consistency.
• Safety, copyright, privacy, and brand risk.

1. Separate objective issues from subjective preferences.
2. Preserve successful elements in any revision advice.
3. Recommend the smallest next action:

• accept as-is.
• local post-process.
• local/AI edit.
• regenerate with revised prompt.
• ask for missing source/reference material.

Return Format

## Image Review

### Verdict
Accept / Minor post-process / Local edit / Regenerate / Need more context

### Scores
- Prompt adherence:
- Technical quality:
- Composition:
- Target-use fit:
- Text/detail reliability:
- Reference consistency:
- Safety/copyright/privacy:

### Issues
1. Issue:
   Evidence:
   Fix:

### Revised Prompt Or Edit Brief
...

### Keep
- Elements that already work:

### Next Step
...

Dependencies

• The image must be visible or readable.
• Prompt/source brief is needed for prompt adherence review.
• Reference images are needed for consistency review.

Limits and Known Issues

• Aesthetic judgment is partly subjective; do not present taste preferences as hard failures.
• Do not give legal clearance for copyright, trademark, publicity rights, or advertising claims.
• Do not provide instructions to bypass model safety systems.
• Do not claim exact generation causes unless they are visible from the image or provided by the user/tool logs.

Examples

User: "这是刚生成的产品海报，能不能发？"

Handling: review title readability, product prominence, crop, artifacts, platform fit, brand risk, and whether local post-process or regeneration is the right next step.

User: "这张图和 prompt 匹配吗？"

Handling: compare subject, style, composition, colors, text, and constraints against the prompt; return mismatches and a revised prompt that keeps the successful elements.