Prompt Composer

Prompt-Composer Skill

Purpose

Eliminates manual prompt construction. Loads pattern library → applies templates → inserts asset specs → outputs Gemini-ready prompt. Prevents copy-paste errors and pattern inconsistency.

Input Format

{
  "asset_id": "ASSET-3",
  "asset_name": "Nocturnal Canopy Pattern",
  "asset_type": "pattern_tile",
  "resolution": "512x512",
  "mode": "Gallery",
  "specimens": ["Eucalyptus", "Wattle", "Banksia", "Gum nuts"],
  "primary_focus": "Botanical only (no fauna)",
  "special_requirements": ["Seamless tiling", "Edge-matching critical"],
  "previous_attempt": {
    "score": 85,
    "violations": ["Upper-left 25% (max 20%)", "Edge mismatch"]
  }
}

Pattern Library Integration

Loads validated patterns from /docs/northcote-asset-generation-patterns.md:

Pattern 1: Production Intent

MAXIMUM RESOLUTION: Generate at highest quality (no compression)
EXPORT FORMAT: PNG with zero artifacts
PRODUCTION ASSET: Not concept exploration—final deliverable

Pattern 2: Geographic Authenticity

ENDEMIC TEST: Every specimen must answer "Did this organism challenge European taxonomy?"
AUSTRALIAN ONLY: NO European fauna, generic specimens, northern hemisphere species
MANDATORY CHECK: [List specimens] → validate against Australian Flora/Fauna Database

Pattern 3: Translucency Tiers

TIER 1 (60-80% transmission): Spider molt, thin membranes
TIER 2 (40-60% transmission): Frill-neck tissue, thick leaves
TIER 3 (0-20% transmission): Structural elements
SYNTAX: "[X-Y% light-transmissive] + [what visible through material]"

Pattern 4: Negative Constraints

For each specimen: "This is NOT [misinterpretation]"
Echidna: NOT full animal body, NOT cute portrait
Spider: NOT live crawling, NOT opaque solid
Frill-neck: NOT standing lizard, NOT profile view

Pattern 5: Density Zone Safety Margins

Upper-left: 15% coverage maximum (NOT 20%) + 200×200px EMPTY
Lower-right: 25% coverage maximum (NOT 30%) + 150×150px EMPTY
Central: 60-80% Wunderkammer density

Pattern 6: Typography Constraints

FONT: Crimson Text or period-appropriate serif
COLOR: Cream #F5F0E8 at 85% opacity
COUNT: 5-6 labels MAXIMUM (reduce for clarity)
FORMAT: "Fig. X. Scientific name (Common name)"

Prompt Assembly Structure

# ASSET [N]: [Name] | [Resolution] | Attempt [N]

## SECTION 1: PRODUCTION INTENT
[Pattern 1 template]

## SECTION 2: VALIDATED PATTERNS FROM PREVIOUS SUCCESSES
From Asset 1 (91/100): [Applicable patterns]
From Asset 2 (94/100): [Applicable patterns]

## SECTION 3: SPECIMEN INVENTORY WITH NEGATIVE CONSTRAINTS
PRIMARY (13-18cm):
- [Specimen name]: [Description] | NOT [misinterpretation]

SECONDARY (8-12cm):
- [Specimen name]: [Description] | NOT [misinterpretation]

TERTIARY (2-8cm):
- [Specimen name]: [Description] | NOT [misinterpretation]

GEOGRAPHIC AUTHENTICITY TEST:
✓ [Specimen]: Challenged European taxonomy via [reason]
✓ [Specimen]: Australian endemic, [taxonomic significance]

## SECTION 4: TRANSLUCENCY PHYSICS (if applicable)
[Only if specimens require transmission]
- [Specimen]: [Tier X] [X-Y% light-transmissive]
  What visible through: [botanical elements, background, other specimens]

## SECTION 5: DENSITY ZONES WITH SAFETY MARGINS
Upper-left: 15% maximum | 200×200px COMPLETELY EMPTY
Lower-right: 25% maximum | 150×150px COMPLETELY EMPTY
Central: 60-80% Wunderkammer organic composition

## SECTION 6: TYPOGRAPHY & DOCUMENTATION
Count: 5-6 labels maximum
Font: Crimson Text serif
Color: Cream #F5F0E8 at 85% opacity
Format: "Fig. 1. Eucalyptus camaldulensis (River Red Gum)"

## SECTION 7: TECHNICAL OUTPUT
Background: #1A1714 (theatrical void, NO sepia drift)
Resolution: [WIDTHxHEIGHT]
Format: PNG maximum quality
[Special requirements: e.g., seamless tiling, edge-matching]

---

## CORRECTIONS FROM ATTEMPT [N-1] (if iteration):
SCORE: [X/100]

WHAT WORKED (preserve):
✓ [Success 1]
✓ [Success 2]

CRITICAL FIXES:
✗ [Violation 1] → FIX: [Specific correction language]
✗ [Violation 2] → FIX: [Specific correction language]

Automation Logic

1. Load asset specifications JSON
2. Query pattern library for applicable patterns
3. If asset_type == "pattern_tile" → add seamless tiling spec
4. If mode == "Laboratory" → apply clinical aesthetic overrides
5. Map specimens → taxonomic significance database
6. Generate negative constraints from common failures
7. If previous_attempt exists → prepend corrections section
8. Format as structured prompt
9. Output ready for Gemini paste

Integration Points

Flash-Sidekick:

• Call consult_pro with specimen list → validate endemic status
• Call analyze_code_quality on draft prompt → identify vague language

Auto-Validator:

• Takes previous attempt's correction_prompt field
• Injects into next generation's CORRECTIONS section

Pattern-Learner:

• Queries pattern library for latest validated patterns
• Applies success learnings from recent high-scoring assets

Token Optimization

Problem: Full prompt ~2000 tokens, Gemini Flash limit 1M tokens/min Solution: Template compression

Remove verbose explanations. Use shorthand:

DENSE (efficient):
"Eucalyptus: 60% transmission, veins visible through leaf tissue"

VERBOSE (wasteful):
"The eucalyptus leaves should demonstrate approximately 60% light transmission, wherein the vein architecture becomes visible when light passes through the leaf tissue, creating an effect of botanical translucency"

Target: 1200-1500 tokens per prompt (25-40% reduction)

Usage Example

prompt = prompt_composer.generate(
    asset_id="ASSET-3",
    asset_name="Nocturnal Canopy Pattern",
    specs={
        "resolution": "512x512",
        "specimens": ["Eucalyptus", "Wattle", "Banksia"],
        "special": ["seamless_tile"]
    },
    previous_attempt={
        "score": 85,
        "violations": ["edge_mismatch", "density_25%"]
    }
)

# Copy-paste to Gemini AI Studio → generate

Efficiency Gain

Before: 10 min manual prompt construction per attempt After: 1 min automated generation Savings: 9 min per attempt × 20-25 attempts = 3-4 hours saved

Error Prevention

• Pattern inconsistency: ELIMINATED
• Copy-paste errors: ELIMINATED
• Missing negative constraints: ELIMINATED
• Forgotten corrections: ELIMINATED
• Token bloat: REDUCED 30%

Transforms pattern library into executable templates. Manual prompt crafting → automated compilation.