Content Engine Dna

content-engine-dna -- Visual DNA Compiler

The Visual DNA Compiler transforms raw visual assets (photos, mood boards, reference videos, brand guidelines, character references) into persistent, structured identity files. Every downstream generation session references these compiled files instead of re-deriving visual intent from scratch.

The core insight comes from two convergent patterns:

1. Karpathy's LLM Wiki -- treat visual identity like source code. Raw assets are the source. The LLM is the compiler. Compiled identity files are the executable. Schema rules enforce consistency. Active linting catches drift.

2. MemPalace spatial hierarchy -- organize compiled knowledge in a navigable structure (brands as wings, characters as rooms, styles as closets). The AAAK 30x compression pattern means a 200-image mood board becomes a 2-page brand DNA file that any tool can consume in a single context load.

The compiler is not a one-shot process. It is iterative: new raw assets trigger recompilation, performance feedback refines compiled prompts, and lint passes catch inconsistencies before they propagate to generated content.

When to Invoke

• User drops reference images, mood boards, or brand assets into knowledge/raw/
• User says "compile brand", "extract DNA", "character sheet", "style guide"
• User says "compile identity" or "brand analysis"
• Before any generation session that needs character consistency or brand alignment
• When compiled files are stale (raw assets modified after last compilation timestamp)
• When /content-engine lint reports provenance errors or orphaned references

Commands

/content-engine compile

Full compilation pipeline. Scans knowledge/raw/ for uncompiled or modified assets, analyzes them via Gemini multimodal, and writes structured identity files to knowledge/compiled/.

Options:

/content-engine compile                        # Full compile (all raw assets)
/content-engine compile --brand {slug}         # Compile specific brand only
/content-engine compile --character {slug}     # Compile specific character only
/content-engine compile --style {slug}         # Compile specific style only
/content-engine compile --incremental          # Only recompile changed raw assets
/content-engine compile --dry-run              # Preview what would be compiled
/content-engine compile --force                # Recompile everything, ignore timestamps

/content-engine lint

Health-check compiled knowledge. Validates provenance, consistency, cross-linking, and tool-specific prompt freshness.

Options:

/content-engine lint                           # Lint all compiled files
/content-engine lint --brand {slug}            # Lint specific brand
/content-engine lint --fix                     # Auto-fix what can be fixed
/content-engine lint --strict                  # Fail on warnings (not just errors)

Compilation Pipeline (6 Steps)

Each step is idempotent. The pipeline can be interrupted and resumed. All intermediate state is written to disk so crashes do not lose work.

Step 1 -- Scan Raw Assets

Traverse knowledge/raw/ and build a manifest of all source files:

knowledge/raw/
├── brand-assets/       # Campaign photos, logos, style guides, mood boards
├── character-refs/     # Face photos, pose references, body type refs
├── style-inspiration/  # Reference reels, screenshots, mood boards
└── scene-briefs/       # Scene descriptions, location photos

For each file, record:

• Path: relative to knowledge/raw/
• Type: image (png/jpg/webp), video (mp4/mov), document (md/pdf), directory
• Hash: SHA-256 of file contents (for change detection)
• Last compiled: timestamp from previous compilation run (if any)

Change detection: Compare current hash against hash stored in the compiled file's sources frontmatter. If different, mark for recompilation. If raw file is new (no corresponding compiled file), mark for initial compilation.

Output: compile-manifest.json (ephemeral, written to .cache/content-engine/).

Step 2 -- Analyze via Gemini Multimodal

For each raw asset marked for compilation, send to Gemini for structured multimodal analysis. The analysis prompt extracts specific visual dimensions (full prompt spec in references/brand-dna-extraction.md).

For images, extract:

• Color palette: Dominant colors (hex values), accent colors, overall temperature (warm/cool/neutral), saturation level
• Composition: Rule of thirds / centered / golden ratio / negative space, subject placement, leading lines
• Lighting: Direction (front/side/rim/overhead), quality (hard/soft/diffused), temperature (kelvin estimate), contrast ratio
• Texture: Surface qualities (matte/glossy/film-grain/digital-clean), depth of field, bokeh character
• Pose language: Subject posture, gesture vocabulary, energy level (candid/editorial/lifestyle/cinematic)

For videos, first extract keyframes via ffmpeg:

ffmpeg -i input.mp4 -vf "select=eq(pict_type\,I)" -vsync vfill -q:v 2 keyframe_%04d.jpg

Then analyze keyframes as images plus:

• Pacing: Cut frequency, shot duration distribution
• Camera movement: Static / dolly / crane / handheld / tracking
• Transitions: Cut / dissolve / whip pan / match cut

API call pattern (using @google/genai):

import { GoogleGenAI } from "@google/genai";

const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });
const response = await ai.models.generateContent({
  model: "gemini-2.0-flash",
  contents: [
    { role: "user", parts: [
      { inlineData: { mimeType: "image/jpeg", data: base64Image } },
      { text: BRAND_DNA_EXTRACTION_PROMPT }
    ]}
  ]
});

The extraction prompt is defined in references/brand-dna-extraction.md. It produces structured JSON that maps directly to the compiled file sections.

Step 3 -- Cross-Reference

Compare new analysis results against existing compiled files:

1. Brand collision detection: If a new brand analysis overlaps >70% with an existing brand DNA (by color palette and composition), flag for merge review rather than creating a duplicate.
2. Character consistency check: If a character reference produces identity traits that conflict with an existing character sheet (e.g., different eye color, different build), flag the conflict and ask the user to resolve.
3. Style deduplication: If a new style analysis matches an existing style guide (by camera language and color grammar), update the existing guide rather than creating a new one.
4. Cross-linking: Identify which characters belong to which brands, which styles apply to which brands, and set up related wikilinks in the compiled files.

Step 4 -- Generate Tool-Specific Prompt Fragments

For each compiled identity, generate exact prompt text that reproduces the identity in each supported tool. This is the critical step that makes compiled knowledge actionable.

Nano Banana Pro prompt fragment:

[Character]: {name}, {age}, {ethnicity}, {build}. {distinguishing features}.
{default expression}. {energy/vibe}.
Wearing: {wardrobe defaults}.
Setting: {environment}.
Lighting: {lighting setup}.

Soul Cinema / Higgsfield prompt fragment:

Cinematic {shot_type} of {subject_description}.
{camera_movement}.
{lighting_direction} {lighting_quality} lighting, {color_temperature}.
{composition_rule}. {depth_of_field}.
Style reference: {director_style}.

Weavy scene prompt fragment:

{scene_description}
Character: {character_sheet_ref}
Environment: {environment_description}
Mood: {mood_keywords}
Lighting: {lighting_description}
Camera: {camera_angle}, {focal_length}

ComfyUI node configuration fragment:

checkpoint: {base_model}
lora:
  name: {lora_name}
  weight: {lora_weight}
sampler: {sampler}
steps: {steps}
cfg: {cfg_scale}
positive: "{positive_prompt}"
negative: "{negative_prompt}"

Each fragment is tested mentally against the tool's known prompt grammar before being written. Anti-patterns (prompts that produce inconsistent results) are documented in the compiled file's Anti-Patterns section.

Step 5 -- Write Compiled Files

Write structured Markdown files to knowledge/compiled/:

knowledge/compiled/
├── brands/{slug}.md          # Brand DNA files
├── characters/{slug}.md      # Character sheets
└── styles/{slug}.md          # Style guides

Each file follows its type-specific format (see Compiled File Formats below). Every compiled file includes:

• YAML frontmatter with provenance metadata
• SHA-256 hashes of all source files in the sources field
• Compilation timestamp
• Tool-specific prompt fragments
• Anti-patterns section

Step 6 -- Verify Provenance

Post-write verification pass:

1. Source integrity: Every sources entry in compiled files points to an existing raw file
2. Hash verification: SHA-256 of each source file matches the hash recorded at compilation time
3. Completeness: Every raw asset directory has at least one corresponding compiled file
4. Cross-link integrity: All [[wikilink]] references in compiled files resolve to existing compiled files
5. Prompt fragment validation: Each tool-specific prompt fragment is syntactically valid for its target tool

Verification failures are logged to .cache/content-engine/compile-report.json and printed to stdout.

Compiled File Formats

Brand DNA (compiled/brands/{slug}.md)

---
name: Brand Name
type: brand-dna
compiled: 2026-04-07T14:30:00Z
sources:
  - path: raw/brand-assets/campaign-photos/hero-01.jpg
    sha256: a1b2c3d4...
  - path: raw/brand-assets/campaign-photos/hero-02.jpg
    sha256: e5f6g7h8...
  - path: raw/brand-assets/logo/primary.svg
    sha256: i9j0k1l2...
tools: [nano-banana-pro, weavy, soul-cinema, comfyui]
related:
  - "[[characters/luna]]"
  - "[[styles/cinematic-warm]]"
status: active
---

# {Brand Name} -- Visual DNA

## Visual Identity

### Color Palette
| Role | Hex | Mood |
|------|-----|------|
| Primary | #2D3436 | Grounding, authority |
| Secondary | #DFE6E9 | Clean, spacious |
| Accent | #E17055 | Warmth, energy |
| Highlight | #FFEAA7 | Optimism, attention |

Temperature: warm-neutral
Saturation: medium (40-60%)

### Composition
- Primary rule: rule of thirds with subject at left intersection
- Negative space: generous (40%+ of frame)
- Leading lines: architectural, converging toward subject
- Depth layering: foreground element (blurred) + subject (sharp) + background (soft)

### Lighting
- Direction: 45-degree key, subtle fill from opposite
- Quality: soft, diffused (overcast or large softbox)
- Temperature: 5200K (warm daylight)
- Contrast ratio: 3:1 (natural, not dramatic)
- Rim light: present but subtle, warm tone

### Texture
- Surface: matte with subtle film grain (ISO 400 emulation)
- Depth of field: shallow (f/2.0-2.8 equivalent)
- Bokeh: smooth, circular
- Digital artifacts: none (clean processing)

### Pose Language
- Energy: relaxed confidence (not stiff, not over-animated)
- Gesture vocabulary: open hands, natural movement
- Eye contact: direct but soft
- Context: lifestyle/editorial hybrid

## Tool-Specific Prompts

### Nano Banana Pro
{exact prompt prefix for this brand's look}

### Soul Cinema
{exact prompt prefix for this brand's look}

### Weavy
{exact prompt prefix for this brand's look}

### ComfyUI
{node configuration for this brand's look}

## Anti-Patterns
- Avoid: harsh direct flash, cool-toned environments, centered symmetrical framing
- Avoid: over-saturated colors (saturation > 70%)
- Avoid: Dutch angles, extreme close-ups without context
- Avoid: pure white backgrounds (use warm off-white #FAF9F6)

Character Sheet (compiled/characters/{slug}.md)

---
name: Character Name
type: character-sheet
compiled: 2026-04-07T14:30:00Z
sources:
  - path: raw/character-refs/luna/face-front.jpg
    sha256: m3n4o5p6...
  - path: raw/character-refs/luna/face-3quarter.jpg
    sha256: q7r8s9t0...
nano_banana_ref: "nb-char-id-12345"
consistency_model: nano-banana-pro
lora_weights: null
face_embedding_hash: "sha256:u1v2w3x4..."
related:
  - "[[brands/acme]]"
  - "[[styles/editorial-warm]]"
status: active
---

# {Character Name} -- Character Sheet

## Identity
- **Age**: 28
- **Ethnicity**: Mixed (Southeast Asian / European)
- **Build**: Athletic-slim
- **Height impression**: Average-tall
- **Hair**: Dark brown, shoulder-length, natural wave
- **Eyes**: Dark brown, almond-shaped
- **Skin tone**: Medium olive (Fitzpatrick IV)
- **Distinguishing features**: Subtle freckles across nose bridge, defined jawline
- **Default expression**: Relaxed confidence, slight asymmetric smile
- **Energy/vibe**: Approachable professional, warm intelligence

## Consistency Anchors
- **Nano Banana Pro character sheet ID**: nb-char-id-12345
- **LoRA weights**: {path or null if using Nano Banana}
- **Face embedding hash**: sha256:u1v2w3x4... (for verification across sessions)
- **Consistency verification**: Compare generated face against reference embeddings; cosine similarity must exceed 0.85

## Scene Defaults

### Wardrobe Palette
- **Core**: Earth tones (olive, rust, cream, slate)
- **Accent**: Warm metallics (gold, bronze)
- **Avoid**: Neon colors, heavy patterns, logos
- **Style**: Smart casual to editorial (structured blazers, quality fabrics)

### Environments That Work
- Urban architecture with natural light
- Cafe/workspace with warm ambient
- Outdoor golden hour with greenery
- Minimalist interiors with texture

### Environments to Avoid
- Sterile office/corporate
- Pure black backgrounds
- Heavily cluttered spaces
- Over-designed sets

### Lighting That Flatters
- 45-degree soft key light (emphasizes cheekbone structure)
- Warm fill (reduces under-eye shadows)
- Subtle rim light (separates from background without halo)
- Avoid: direct overhead (harsh nose shadow), direct front (flattens features)

Style Guide (compiled/styles/{slug}.md)

---
name: Style Name
type: style-guide
compiled: 2026-04-07T14:30:00Z
sources:
  - path: raw/style-inspiration/wes-anderson-refs/
    sha256: y5z6a7b8...
category: cinematic
camera_ref: [wes-anderson]
lora_ref: "wes-anderson-pastel-v2.safetensors"
related:
  - "[[brands/acme]]"
  - "[[characters/luna]]"
status: active
---

# {Style Name} -- Style Guide

## Camera Language
- **Shot types**: Medium-wide (primary), centered close-up (accent), symmetrical wide (establishing)
- **Movement**: Static (primary), slow lateral dolly (accent), centered zoom-out reveal
- **Framing**: One-point perspective, perfect symmetry, centered subjects
- **Focal length**: 35mm-50mm equivalent (no wide-angle distortion, no telephoto compression)

## Color Grammar
- **Primary palette**: Pastel (muted pink, powder blue, butter yellow, sage green)
- **Secondary**: Warm neutrals (cream, camel, terracotta)
- **Grading**: Desaturated 15% from capture, raised shadows, compressed highlights
- **LUT reference**: Wes Anderson Pastel Pack #3 (or manual: lift shadows +10, lower highlights -5, desaturate 15%)

## Lighting Setup
- **Key**: Large diffused source, frontal to 30-degree offset
- **Fill**: Bounce from opposite side, 2:1 ratio (very low contrast)
- **Rim**: Warm practical lights in frame (lamps, windows)
- **Temperature**: 5600K overall, warm practicals at 3200K for depth
- **Quality**: Extremely soft, almost shadowless on face

## Exact Prompts

### Nano Banana Pro
{prompt fragment}

### Soul Cinema
{prompt fragment}

### Weavy
{prompt fragment}

### ComfyUI
{node configuration with LoRA reference}

Linting Rules

The lint pass enforces the Karpathy active-maintenance pattern: compiled knowledge must stay consistent with its sources and with itself.

Consistency Rules

Orphaned Reference Rules

Cross-Linking Rules

Lint Output

{
  "timestamp": "2026-04-07T14:30:00Z",
  "files_checked": 12,
  "errors": 0,
  "warnings": 3,
  "details": [
    {
      "file": "compiled/brands/acme.md",
      "rule": "stale-compiled",
      "severity": "warning",
      "message": "Source raw/brand-assets/campaign-photos/hero-03.jpg modified after last compilation"
    }
  ]
}

Errors block the pipeline (generation should not proceed with broken identity files). Warnings are informational and suggest corrective action.

Dependencies

Required

Optional

File Locations

Self-Maintenance Rules

These rules govern any agent that modifies files in this skill. They are enforced by reasoning, not by hooks.

Rule 1 -- Pipeline step count consistency When adding or removing compilation steps: update the step count in BOTH this file's heading ("6 Steps") AND any references to the pipeline in sibling SKILL.md files. The pipeline step count must always match across all files.

Rule 2 -- Compiled file format consistency When adding a new compiled file type or modifying frontmatter fields: update BOTH the format examples in this file AND the corresponding template in templates/. The template must always reflect the current schema.

Rule 3 -- Linting rule consistency When adding a new lint rule: add it to the Linting Rules table in this file AND implement the check in scripts/compile-dna.py. Rules documented but not implemented are worse than no rules.

Rule 4 -- Tool-specific prompt format consistency When adding a new supported tool: add prompt fragment format examples in this file, in references/brand-dna-extraction.md (extraction prompt), and in references/style-locking.md (if the tool uses LoRA or equivalent). All three files must agree on supported tools.

Rule 5 -- SKILL.md is authoritative This SKILL.md is the single source of truth for compilation pipeline, compiled file formats, and linting rules. All reference files defer to it. If a conflict exists, this file wins.

Integration Points