# Production Director Benchmark v1

> Skill domain: AI production director / prompt engineering for image generation
> Scoring scale: 1–10 per dimension
> Weighted total = 100%
> Target benchmark line: ≥95

---

## Scoring Dimensions

### 1. Requirement Parsing (20%)
Ability to extract structured intent from a raw natural-language demand.

| Score | Description |
|-------|-------------|
| 1–3   | Misses core subject, style, or exclusions; output is substantially off-brief. |
| 4–5   | Captures obvious elements but omits nuanced details (lighting, mood, aspect-ratio needs, specific no-gos). |
| 6–7   | Parses subject, render style, key details, and explicit exclusions with minor gaps in implicit constraints. |
| 8–9   | Precise extraction of all visible entities, moods, dimensions, format needs, layout rules, and branding risks. |
| 10    | Perfect parsing: captures explicit and implicit requirements, downstream-use constraints (e.g. text overlay zones), and edge-case exclusions. |

### 2. Engine Selection (15%)
Chooses the optimal generative engine for the scene style, text needs, consistency, and production constraints.

| Score | Description |
|-------|-------------|
| 1–3   | Mismatched engine for the task (e.g. DALL·E 3 for text-heavy logo, Midjourney for precise typography). |
| 4–5   | Engine works but is clearly sub-optimal; better alternatives exist for cost, control, or quality. |
| 6–7   | Reasonable engine choice aligned with common heuristics (MJ for aesthetics, SDXL for character consistency, Flux for text rendering). |
| 8–9   | Strong choice justified by style, controllability, text-rendering requirements, and batch-consistency needs. |
| 10    | Optimal engine with implicit justification: leverages the engine’s unique strength (e.g. Ideogram for embedded text, Flux.1 for ad copy, ComfyUI for ControlNet). |

### 3. Prompt Quality — Positive (20%)
Clarity, structure, vividness, and engine-native phrasing of the positive prompt.

| Score | Description |
|-------|-------------|
| 1–3   | Sparse, vague, or grammatically broken prompt; little scene coherence. |
| 4–5   | Covers basics but lacks structure, quality enhancers, or layering (subject → details → environment → lighting → style → quality). |
| 6–7   | Structured prompt with clear subject, style, lighting, and some quality enhancers; minor engine-syntax issues. |
| 8–9   | Rich, layered prompt following best-practice order; strong vocabulary, correct engine syntax, and artifact-aware phrasing. |
| 10    | Master-level prompt: vivid, precise, ordered, includes engine-native flags, uses appropriate sentence vs. keyword style, and weaves in negative-prompt synergies. |

### 4. Prompt Quality — Negative / Exclusion Guard (15%)
Completeness and relevance of the negative prompt mapped to the requirement no-gos and engine weaknesses.

| Score | Description |
|-------|-------------|
| 1–3   | No negative prompt or completely irrelevant exclusions. |
| 4–5   | Generic negative prompt that partially overlaps with requirements; misses scene-specific no-gos. |
| 6–7   | Covers common artifacts (watermark, text, blurry, low quality) and some explicit exclusions from the demand. |
| 8–9   | Comprehensive negative list mirroring exclusions; tuned to target engine weaknesses (hands for SD/MJ, text errors for non-Flux, realism for stylized art). |
| 10    | Surgical negative prompt: maps every exclusion, adds engine-specific anti-patterns (anatomy locks, style locks, environment locks), and includes quality gates without over-filtering. |

### 5. Parameter & Specification Accuracy (20%)
Correctness and consistency of dimensions, aspect ratio, engine parameters, output format, and resolution.

| Score | Description |
|-------|-------------|
| 1–3   | Wrong aspect ratio, missing dimensions, invalid engine parameters, or format/resolution mismatch. |
| 4–5   | Parameters mostly present but inconsistent (e.g. dimension mismatch with stated ratio, outdated or mixed engine flags). |
| 6–7   | Correct aspect ratio, resolution, and basic engine flags; minor inconsistencies in seed, sampler, or output count. |
| 8–9   | Fully consistent spec: dimensions match ratio, engine-native parameters filled correctly, output format and resolution appropriate for the use case. |
| 10    | Flawless spec: dimensions, ratio, engine parameters (including advanced flags like --stylize, sref, ControlNet, LoRA), reference-image handling, output format/resolution, and workflow placeholders all production-ready. |

### 6. Domain Safety & Layout Awareness (10%)
Risk management for text overlays, brand safety, political sensitivity, NSFW, and downstream UI/layout constraints.

| Score | Description |
|-------|-------------|
| 1–3   | Generates NSFW-adjacent, politically sensitive, or copyright-risk content without mitigation; ignores layout needs. |
| 4–5   | Vague awareness of issues but no concrete guidance in quality_notes or parameter choices. |
| 6–7   | Notes basic留白/水印 avoidance and generic safety cues; some layout guidance present. |
| 8–9   | Explicitly flags text-overlay zones, logo replacement areas, border/politics sensitivity, and children-safe content in quality_notes. |
| 10    | Proactive risk management: quantified留白 guidance, logo/brand placeholder strategy, country-border abstraction, age-appropriate guardrails, and post-processing hints documented clearly. |

---

## Weighted Total Formula

```
Total = 0.20 × Requirement Parsing
      + 0.15 × Engine Selection
      + 0.20 × Prompt Quality (Positive)
      + 0.15 × Prompt Quality (Negative)
      + 0.20 × Parameter & Specification Accuracy
      + 0.10 × Domain Safety & Layout Awareness
```

**Target benchmark line: ≥95**