AI Art Directors Don’t Type ‘Make It Pretty’

📖 6 min read

The “Make It Look Good” Prompt Is Why Your Visuals Look AI-Generated

You type “futuristic city, cinematic lighting, 8K, masterpiece” into Midjourney. You get something that looks impressive for 3 seconds and generic forever after. It could be anyone’s image. It communicates nothing specific. It has no point of view.

That’s because you’re using AI as a slot machine instead of as a production tool.

Art directors at agencies billing $200/hour for AI-assisted work don’t type descriptions and hope. They run a layered visual process – the same way a photography shoot has pre-production, shooting, and post. The AI is the camera. You still need the director.

Why Single Prompts Produce “AI Art”

The tell-tale signs of one-prompt generation:

• Technically impressive but emotionally empty
• Over-rendered (everything is detailed, nothing is focal)
• Disconnected from any brand or visual language
• Looks like 10,000 other images generated that day
• No compositional intention – the AI just filled the frame

The problem isn’t the tool. It’s that you’re making the AI decide everything simultaneously – mood, composition, color, style, detail level, and meaning – in one instruction. That’s like telling a photographer “take a good photo” without a brief.

The 5-Layer Visual Pipeline

Layer 1: Mood Board Construction

Before generating anything, define the emotional and tonal direction.

The prompt:

I need a visual for [purpose/context]. The audience is [who] and it needs to communicate [core message/emotion].

Don't generate an image yet. Instead, build me a mood board brief:

1. Emotional tone: What should the viewer FEEL? (list 3 specific emotions, ranked)
2. Reference universe: What existing visual worlds does this live in? (films, photographers, art movements, brands)
3. Color psychology: What palette communicates this emotion? (specific hex ranges, not just "blue")
4. Texture and materiality: Smooth/rough, organic/synthetic, aged/pristine?
5. Light quality: What kind of light tells this story? (time of day, direction, quality, color temperature)
6. Anti-references: What should this explicitly NOT look like? (equally important)

Be specific enough that two different designers would produce similar results from this brief.

Why this works: The anti-references are crucial. They prevent the AI from defaulting to the most common aesthetic in its training data. “NOT like a stock photo” or “NOT the cyberpunk cliche” gives the model genuine constraints to push against.

Layer 2: Composition Framework

Decide where everything goes before you decide what it looks like.

The prompt:

Based on this mood direction:
[paste Layer 1 output]

Design the composition (no generation yet):

1. Aspect ratio and format: [specify based on use - social, banner, editorial, etc.]
2. Focal hierarchy: What does the eye hit first, second, third?
3. Rule of thirds / golden ratio / centered - which composition principle and why?
4. Negative space: Where is it and what purpose does it serve? (breathing room, text placement, tension)
5. Depth layers: Foreground, midground, background - what occupies each?
6. Scale relationships: What's big vs. small and what does that communicate?

Draw this as a text-based wireframe if possible. Mark zones as [FOCAL], [SECONDARY], [BACKGROUND], [NEGATIVE SPACE].

Why this works: Composition is where meaning lives. A subject in the center communicates something different than the same subject in the lower third with vast sky above. By deciding this before generation, you create intentional images instead of AI’s default “centered subject, blurred background.”

Layer 3: Style-Specific Generation

NOW you generate – with all the decisions already made.

The prompt (for your image generation tool):

[Compose your generation prompt using:]

Subject: [from Layer 2 focal hierarchy]
Composition: [from Layer 2 framework]
Mood: [from Layer 1 emotional tone]
Lighting: [from Layer 1 light quality]
Color: [from Layer 1 palette]
Style reference: [from Layer 1 reference universe]
NOT: [from Layer 1 anti-references]

[Platform-specific parameters: --ar, --style, --stylize, etc.]

Then evaluate what you get:

Look at these generated results against my brief:
[describe or reference the outputs]

Score each against:
- Emotional accuracy (does it evoke the right feelings?) /10
- Compositional accuracy (does it follow the framework?) /10
- Style consistency (does it match the reference universe?) /10
- Originality (does it avoid the anti-references?) /10

Which output best matches? What specific adjustments would improve the next round?

Why this works: You’re not hoping for the right image. You’re evaluating against clear criteria. This makes iteration purposeful instead of random re-rolling.

Layer 4: Detail and Refinement Pass

Take your best output and refine specific areas.

The prompt:

This image is 80% there. Here's what needs refinement:
[describe the image and what's working]

Fix these specific issues:
1. [Specific area] needs [specific change] because [reason tied to brief]
2. [Detail] is [too much/too little] - adjust to [specific level]
3. [Element] is pulling attention from [focal point] - reduce its visual weight

Maintain: [list what's working that must not change]

Generate refined versions focusing on these adjustments only.

For AI tools with inpainting or regional editing, this becomes surgical. For text-based iteration, describe what to keep and what to change.

Why this works: Refinement with clear instructions prevents the common failure of “try again” generating something completely different. You protect what works while fixing what doesn’t.

Layer 5: Brand Audit

Final check against brand standards and use context.

The prompt:

Audit this final visual against these criteria:

Brand standards:
- Does it match our visual language? [describe brand aesthetic]
- Is it consistent with our other recent visuals?
- Does it work at all required sizes? (thumbnail, full-size, cropped)

Technical check:
- Is there space for text overlay where needed?
- Are there any AI artifacts (weird hands, impossible geometry, text gibberish)?
- Does the focal point survive mobile crop?

Context check:
- Will this stand out in [specific feed/context where it'll appear]?
- Could this be confused with a competitor's visual?
- Does it pass the "would I stop scrolling" test at thumbnail size?

Pass/fail each criterion. If any fail, specify the fix.

Why this works: This catches the gap between “looks cool” and “works for the actual use case.” An image that’s gorgeous at full size but loses its meaning at 400×400 is useless for social.

The Production Difference

One-prompt results:

• Takes 2 minutes, looks like everyone else’s AI art
• No clear focal point or compositional intention
• Disconnected from brand or message
• Gets replaced by the next generic generation tomorrow

Layer Method results:

• Takes 30-45 minutes for a hero image
• Has clear visual intention and emotional direction
• Works within brand system
• Could pass as art-directed photography or illustration
• Has longevity because it communicates something specific

When to Use Each Layer

Full 5 layers: Hero images, campaign visuals, brand assets, editorial illustrations, anything client-facing at scale.

Layers 1 + 3 only: Social content, blog headers, internal presentations – where good-enough beats perfect.

Layer 3 only: Brainstorming and exploration – when you’re finding direction, not producing finals.

Copy This Workflow