creative-eye

What this skill does

# Creative Eye — Design Judgment for AI Agents

## The Problem

AI agents ship bad creative because they have no taste. Specifically:

1. **They confuse "exists" with "good."** Generating output ≠ creating quality. The first version is almost never good enough, but agents treat it as done.
2. **They have no benchmarks.** Without seeing what "great" looks like, agents can't distinguish amateur from professional work.
3. **They optimize for volume over quality.** 50 mediocre variations when 1 excellent piece would outperform all of them combined.
4. **They can't self-evaluate.** They lack the vocabulary and frameworks to critique their own work.
5. **They don't learn from failures.** Same mistakes repeat across sessions because there's no feedback loop.

This skill provides the frameworks, prompts, and workflows to fix all five problems.

---

## The Framework: STUDY → COMPARE → CREATE → EVALUATE

Every creative task follows this sequence. Never skip steps.

### 1. STUDY — Build Taste Through Exposure

Before creating anything in a given domain, study what "great" looks like.

**Daily practice (10 min):**
Pick ONE brand or creator. Analyze ONE piece of their content. Ask:
- What makes it work visually? Be specific (not "it looks good" but "the 200-tracking on uppercase Didot creates editorial authority")
- What are the exact typography choices? (font, weight, size, tracking, color)
- What's the composition doing? (where does the eye land first, second, third?)
- What role does negative space play?
- What would break if you changed one element?
- What feeling does it create, and HOW does it create that feeling?

Log findings to a study log file. Accumulate observations over time. This builds the vocabulary and pattern library that enables judgment.

**Picking brands to study:**
Choose 8-12 brands across these tiers:
- **Tier 1 (3-4 brands):** Direct competitors or brands with a similar aesthetic to yours
- **Tier 2 (3-4 brands):** Brands your target audience admires, even if different category
- **Tier 3 (2-3 brands):** Category leaders with exceptional design systems
- **Tier 4 (1-2 brands):** Wildcard inspiration from unrelated fields (architecture, editorial, fashion)

### 2. COMPARE — Find References Before Creating

Before generating any visual content:
1. Find 3-5 reference examples of what "great" looks like for THIS specific format
2. Save them or note URLs
3. Identify the specific qualities that make each one great
4. Use those qualities as the creative brief constraints

**Never create in a vacuum.** The difference between amateur and professional creative is almost always that professionals looked at references first.

### 3. CREATE — Every Decision Needs a Reason

When generating creative, every choice must be intentional:
- **Why this font?** Not "it's available" — what does it communicate? (Serif = editorial authority. Geometric sans = modern tech. Humanist sans = friendly approachable.)
- **Why this color?** Does it match the brand palette? What emotion does it carry?
- **Why this layout?** What does left-aligned vs centered communicate? Where should the eye go first?
- **Why this size?** Small and restrained = premium/exclusive. Large and bold = loud/promotional.
- **Why this image treatment?** Warm and grainy = authentic/vintage. Clean and sharp = modern/clinical.

If you can't articulate why, you're guessing. Stop and refer back to your references.

### 4. EVALUATE — Score Before Shipping

Run every piece through both evaluation tools below before publishing.

---

## The 5-Point Creative Scorecard

Adapted from Runway's video evaluation framework. Score each dimension 1-10.

| # | Dimension | Question | Min Score |
|---|-----------|----------|-----------|
| 1 | **Brief Adherence** | Does this serve the stated business/creative goal? | 7 |
| 2 | **Brand Consistency** | Does this look like it came from the same brand as everything else? | 8 |
| 3 | **Visual Quality** | No artifacts, misspellings, clip-art vibes, AI tells? | 9 |
| 4 | **Emotional Resonance** | Would a real human stop scrolling for this? | 7 |
| 5 | **Style Match** | Does this match the brand's specific aesthetic? | 8 |

**If ANY score is below its minimum → do not publish. Fix or discard.**

---

## The 10-Point Pre-Publish Checklist

Quick yes/no gate. Requires 8+ "yes" to ship.

1. Does this look like it came from a well-funded brand? (not a weekend side project)
2. Would your most design-savvy friend say "this is fire" unprompted?
3. If you remove all text, is the composition still interesting?
4. Does the typography have personality, or is it just "text on a thing"?
5. Are there more than 3 elements competing for attention? (if yes, remove one)
6. Would someone in your target audience actually use/wear/share this?
7. Is there anything that looks AI-generated or clip-art-like?
8. Have you compared this side-by-side with a real brand reference?
9. Would you personally pay the listed price for this?
10. If this got 100K views, would it help or hurt the brand?

---

## The Self-Refine Loop

Based on Andrew Ng's Reflection pattern and the Self-Refine research framework (CMU). This is the core workflow for iterative improvement.

```
Generate → Critique (vision model) → Fix → Critique again → Ship or Kill
```

### The Loop

**Step 1: Generate.** Create the first draft using references and brand constraints.

**Step 2: Critique.** Use a vision model to evaluate the output. Feed it the generated image plus 2-3 reference images from your library. Use the evaluation prompts below. Get specific scores and specific issues.

**Step 3: Fix.** Address ONLY the specific issues identified. Do not regenerate from scratch unless the critique identifies fundamental problems (wrong concept, wrong style direction).

**Step 4: Re-critique.** Run the improved version through the same evaluation. Compare scores.

**Step 5: Ship or Kill.**
- All scores at or above minimums → ship it
- Improved but still below on 1-2 dimensions → one more iteration (go to Step 3)
- No meaningful improvement after iteration → kill it and try a different approach
- **Maximum 3 iterations.** If it's not working after 3 passes, the concept is wrong, not the execution. Escalate to a human or start over with a fundamentally different direction.

### Why 3 Iterations Max

Diminishing returns hit fast. If the core concept is sound, 1-2 refinements will get it there. If it takes more than 3, you're polishing the wrong thing. The discipline to kill bad work is as important as the ability to refine good work.

---

## Vision Model Evaluation Prompts

Use these with any vision-capable model (GPT-4o, Claude, Gemini). Feed the generated image alongside 2-3 reference images from aspirational brands.

### Merch Design Review

```
You are a senior creative director at a premium lifestyle brand.

Review this merch design against these criteria:

1. TYPOGRAPHY: Is the type treatment sophisticated? Does it have personality and intentional styling (tracking, weight, case)? Or does it look like default text on a template?
2. COMPOSITION: Is the layout balanced and intentional? Does whitespace serve the design? Or is it cluttered or awkwardly empty?
3. BRAND ALIGNMENT: Does this feel like it belongs to a specific brand with a clear identity? Or is it generic?
4. WEARABILITY/USABILITY: Would someone in the target demographic actually buy and use this? Would they be proud to be seen with it?
5. PRODUCTION QUALITY: Does this look like it was designed by a professional? Or does it have AI-generation artifacts, clip-art qualities, or amateur touches?

Score each dimension 1-10. Be brutally honest.
For any score below 7, explain exactly what's wrong and provide a specific fix.
Compare against the reference images provided — how does this stack up?
```

### Social Content Review

```
You are a social media creative director for a premium brand.

Review this social post (image + copy) against:

1. SCROLL-STOPPING POWER: Would this make s