---
name: ck-ai-artist
description: >
  Generates, edits, and refines images using AI image generation APIs.
  Activate when user says 'generate an image', 'create artwork', 'AI art',
  'draw me', 'make a picture of', or 'image generation prompt'. Accepts
  text descriptions, style references, dimensions, and iteration requests.
---

## Overview
Crafts optimized prompts and orchestrates AI image generation workflows. Handles prompt engineering, style specification, negative prompts, and iterative refinement for high-quality image outputs.

## When to Use
- User requests visual content creation from text descriptions
- Generating concept art, illustrations, or mockup visuals
- Iterating on existing images with style or composition changes
- Batch generating image variants for A/B testing or design exploration
- Creating consistent image sets with shared style parameters

## Don't Use When
- User needs photo editing of an existing uploaded image (use ck-ai-multimodal)
- Task requires vector graphics or SVG creation (use a design tool)
- User needs video generation (use ck-remotion or a video AI service)
- The output will be used in a commercial context without verifying model license

## Steps / Instructions

### 1. Clarify Image Requirements
Gather before generating:
- Subject / main content
- Style (photorealistic, illustration, watercolor, pixel art, etc.)
- Mood / lighting / color palette
- Dimensions / aspect ratio (e.g., 16:9, square, portrait)
- Reference styles or artists (check license implications)

### 2. Craft the Prompt

**Structure:**
```
[subject], [environment/setting], [style], [lighting], [camera/perspective], [quality modifiers]
```

**Example:**
```
A futuristic city skyline at dusk, neon reflections on wet streets,
cyberpunk illustration style, cinematic lighting, wide angle,
high detail, vibrant colors
```

**Negative prompt (things to avoid):**
```
blurry, low quality, text, watermark, distorted faces, extra limbs
```

### 3. Select Model / API Parameters

Common parameters across providers:
```json
{
  "prompt": "<optimized prompt>",
  "negative_prompt": "<what to avoid>",
  "width": 1024,
  "height": 1024,
  "steps": 30,
  "guidance_scale": 7.5,
  "seed": 42
}
```

Provider options (do not hardcode API keys — use env vars):
- `OPENAI_API_KEY` → DALL-E 3
- `STABILITY_API_KEY` → Stable Diffusion / SDXL
- `REPLICATE_API_TOKEN` → various open models

### 4. Iterate and Refine

After first generation:
1. Review output against intent
2. Adjust prompt — add specifics for weak areas
3. Tweak guidance scale (higher = more prompt-adherent, less creative)
4. Try seed variations for composition changes
5. Use img2img if base composition is close but needs refinement

### 5. Post-Processing Suggestions
- Upscale with Real-ESRGAN for print-quality output
- Use inpainting to fix specific regions
- Apply style transfer for consistent series

## Prompt Engineering Tips
- Be specific about count: "three red apples" not "some apples"
- Lighting has major impact: "golden hour", "studio softbox", "moonlight"
- Add medium: "oil painting", "digital illustration", "pencil sketch"
- Quality boosters: "highly detailed", "sharp focus", "8k", "award-winning"
- Avoid abstract adjectives alone: "beautiful" is weak; describe what makes it beautiful

## Notes
- Never embed API keys in scripts; always read from environment variables
- Log prompt + seed + parameters for reproducibility
- Respect content policies of each image generation provider