| name | Visual Creation |
| description | Quality-gated image generation workflow. Enforces 6-step pipeline with council review, organic-first principle, and human approval gate before batch generation. Use for ALL image generation tasks. |
Visual Creation Skill
Purpose
Structurally enforced quality pipeline for image generation. Every step blocks the next — you cannot skip council review, you cannot batch without human approval, you cannot generate without studying the style DNA first.
This skill supersedes direct calls to Nano Banana MCP for any brand-visible image.
The 6-Step Pipeline
Step 1: STUDY (Required — Cannot Skip)
Before drafting any prompt, read and internalize:
- Read
docs/strategy/IMAGE_STYLE_DNA.md — the canonical style reference
- Identify which of the 3 approved styles fits:
- Dark Premium Glass (70% — architecture, ecosystem, products)
- Da Vinci Codex (15% — research, complete guides)
- Light Workshop (15% — tutorials, workshops)
- Select sub-variant if Dark Premium Glass:
- C1 Cyberpunk Neon (comparisons, roadmaps)
- C2 Da Vinci Dark (ecosystem, flagship)
- C3 Cinematic Studio (music, creative)
- C4 Holographic Command (AI orchestration)
- Choose organizing metaphor: A physical object or scene that structures the composition (armillary sphere, inventor's workbench, neural forest canopy, crystal cave laboratory). NOT "a futuristic visualization."
- Name 4+ components: Each gets a CAPS LABEL (THE SKILL ENGINE, THE MEMORY VAULT). No "a gear" or "some icons."
Self-check before proceeding: Can you name the style, sub-variant, metaphor, and 4+ components? If not, you haven't studied enough.
Step 2: DRAFT (Structured Template Required)
Load the appropriate template from data/visual-prompt-templates.json and fill ALL fields. Every field is mandatory — empty fields block progression.
Template anatomy (all fields required):
SUBJECT: [What this image depicts — specific, not generic]
ORGANIZING METAPHOR: [Physical object/scene structuring the composition]
STYLE: [Approved style name + sub-variant]
VIEWPOINT: [Camera angle — isometric, bird's eye, dramatic low angle, etc.]
COMPONENTS (minimum 4):
1. NAME: [CAPS LABEL] | POSITION: [where in frame] | COLOR: [hex] | MATERIAL: [tactile description]
2. NAME: [CAPS LABEL] | POSITION: [where in frame] | COLOR: [hex] | MATERIAL: [tactile description]
3. NAME: [CAPS LABEL] | POSITION: [where in frame] | COLOR: [hex] | MATERIAL: [tactile description]
4. NAME: [CAPS LABEL] | POSITION: [where in frame] | COLOR: [hex] | MATERIAL: [tactile description]
ORGANIC ANCHOR: [Living/organic element — creature, plant, texture, growth pattern]
FOREGROUND: [What's closest to viewer]
MIDGROUND: [Central subject area]
BACKGROUND: [Atmospheric depth]
ATMOSPHERE: [Lighting, mood, particles, fog, glow]
EMOTIONAL TONE: [What viewer should feel]
NEGATIVE: [What to explicitly exclude — always include standard anti-patterns]
Organic Anchor rule: Every prompt MUST include an organic anchor. This is the single most critical field. Reference: mascot-v05-techno-beast-standing.png — fur before circuits, character in eyes, glow from within.
Step 3: COUNCIL (Blocking Agent Call)
Submit the drafted prompt to the visual-creation-council agent for review.
This step is BLOCKING. You cannot proceed without an APPROVED verdict.
Invoke: visual-creation-council agent
Input: The completed template from Step 2
Expected output: Score + verdict (APPROVED / NEEDS-REVISION / REJECTED)
If NEEDS-REVISION: Address ALL feedback points, resubmit. Maximum 3 revision cycles.
If REJECTED: Fundamentally rework the prompt. Do not iterate on a rejected concept.
If APPROVED: Proceed to Step 4.
Step 3.5: OPTIMIZE PROMPT (Two-Model Pattern)
Before sending to the image model, optimize the prompt using a text model:
- Take the approved template from Step 3
- Use Claude/Gemini Flash to expand it with:
- Missing visual details (lighting direction, camera lens, time of day)
- Material textures (brushed metal, leather grain, frosted glass)
- Composition guidance (rule of thirds, leading lines, depth layers)
- Negative prompt enhancement (add common failure patterns to avoid)
- The text model's job: fill gaps the human didn't think of, NOT change the creative direction
Why this works: Image models produce dramatically better results when prompts include specific physical details (lighting, texture, camera) rather than abstract concepts. A text model bridges the gap between "what you mean" and "what the image model needs to hear." Research shows prompt scores go from ~49/100 to 95/100 with this optimization step.
Skip for: Simple test generations or when the prompt is already highly detailed.
Step 4: TEST (Human Gate — Blocking)
Generate exactly 1 image using the approved prompt.
mcp__nanobanana__generate_image({
prompt: "{approved_prompt}",
aspect_ratio: "16:9",
model_tier: "pro",
enable_grounding: true,
thinking_level: "high",
resolution: "high",
});
Show the test image to the user. Wait for explicit feedback.
This is the core fix for the "24 chrome robots at once" problem. One image. Human eyes. Then decide.
Do NOT proceed to Step 5 or 6 until the user responds.
Acceptable user responses:
- "Looks good" / "Approved" → Proceed to Step 6 (or skip if only 1 image needed)
- "Change X" → Proceed to Step 5 (ITERATE)
- "Start over" → Return to Step 2
Step 5: ITERATE (If Needed)
Apply user feedback to the prompt. If changes are significant (new metaphor, different style, major composition shift), re-submit to council (Step 3). If changes are minor (color tweak, label adjustment), proceed directly.
Maximum 2 iteration cycles in this step. If still not satisfactory, return to Step 2 with fresh approach.
Step 6: EXECUTE (Controlled Batch)
Generate a batch of 3-5 images maximum (not 24).
For each generated image, apply the 7-gate quality filter:
| Gate | Check |
|---|
| 1. Brand Alignment | Matches approved style? |
| 2. Color Balance | 3+ accent colors visible? |
| 3. Information Density | Every area purposeful? |
| 4. Depth Layers | Foreground/mid/background identifiable? |
| 5. Text Legibility | Labels readable at 680px? |
| 6. Icon Quality | Rich 3D, not flat clipart? |
| 7. Scroll-Stop | Would you pause scrolling? |
Plus the organic warmth test: Does this image have living presence, or is it sterile chrome?
Score each image 1-10. Flag any scoring < 6 for regeneration.
Log results to data/image-generation-log.json.
Quick Reference: What Makes Prompts Fail
| Failure Pattern | Why It Fails | Fix |
|---|
| "A visualization of AI systems" | No specificity, no metaphor | Name a physical metaphor: "An inventor's clockwork observatory" |
| Components listed as "gears, icons, nodes" | Generic, unnamed, no color | "THE MEMORY VAULT — bottom-left — chrome and obsidian — #43BFE3 glow" |
| No organic element | Produces sterile chrome renders | Add creature, root network, crystal growth, bioluminescent moss |
| Single color palette | Monotone, boring | Always specify 3+ hex codes from brand palette |
| "Futuristic" as sole style direction | Means nothing to image model | Specify exact surfaces: "glass panels with chrome bezels, navy background" |
Integration
- Council agent:
.claude/agents/visual-creation-council.md
- Style reference:
docs/strategy/IMAGE_STYLE_DNA.md
- Templates:
data/visual-prompt-templates.json
- Generation log:
data/image-generation-log.json
- Regeneration queue:
data/image-regeneration-queue.json
- MCP: Nano Banana (Gemini 2.5 Flash Image)
- Skill rules:
.claude/skills/skill-rules.json (triggers on image generation keywords)
When NOT To Use This Skill
- Quick throwaway test generations (not for production/brand use)
- Screenshots or screen captures
- User provides their own complete prompt and explicitly opts out of council review
For all brand-visible images — blog heroes, social posts, product visuals, mascot scenes — this skill is mandatory.
Visual Creation Skill v1.0 — Feb 21, 2026
Quality over quantity. One excellent image beats twenty mediocre ones.
