Art

Art Skill

Complete visual content system supporting multiple brand aesthetics.

When Not to Use

• AI-generated photoreal images or video -- use comfyui (local diffusion) or clipcannon (video) instead
• Academic/publication diagrams destined for a LaTeX document -- prefer TikZ via book-publishing (references/latex-conventions.md); only reach for this skill's diagram-upcycling.md path to polish a TikZ/Mermaid render, not to author the diagram itself
• Plain image editing (resize, crop, format convert, batch watermark) with no generation involved -- use imagemagick

Aesthetic Routing

The art skill supports multiple visual identities. Before generating any visual, determine which aesthetic applies.

How to Select an Aesthetic

1. Check if the user specifies a brand/project (e.g., "my-brand icon", "acme newsletter header")
2. Check for content-type exceptions (see Content-Type Override Rules below)
3. Check if the target output has a site-specific style guide (e.g., HERO-IMAGE-TEMPLATE.md)
4. If no brand specified, use the default aesthetic

Content-Type Override Rules

CRITICAL: Some content types intentionally use different aesthetics than their host site.

Rule: When content-type override detected, DO NOT check host site style guides. The content is intentionally divergent.

Base Prompt Prefix Standard (MANDATORY)

Every aesthetic MUST define a Base Prompt Prefix — a locked consistency block that gets prepended to every image generation prompt for that aesthetic. This ensures visual cohesion across a set of illustrations.

What the prefix locks down:

• Line weight and quality
• Background treatment
• Camera angle / perspective
• Fill ratio (illustration vs. whitespace)
• Color balance percentages
• Shadow intensity and direction
• Text treatment (if applicable)

Where to find each prefix:

• Each aesthetic file (.md) contains a "Base Prompt Prefix" section with the locked values
• Project-specific aesthetics (e.g., lead magnet AESTHETIC.md) inherit from their parent brand and override only what changes
• Always read the aesthetic file and use its prefix before generating — never freestyle prompt parameters that have been locked

When creating new aesthetics via aesthetic-definer: The agent MUST produce a base prompt prefix as part of the output. An aesthetic without a consistency lock is incomplete.

Available Aesthetics

To add a new aesthetic: Create a new file at aesthetics/[name].md using the default aesthetic as a template. See aesthetics/README.md for the required format.

Aesthetic Loading Rule

Read the matching aesthetic file BEFORE writing any prompt or generating any image. The aesthetic file is the source of truth for that brand's colors, style parameters, anti-patterns, and prompt integration phrases.

Workflow Routing

Image Generation

Default model: nano-banana-2 (Gemini 3.1 Flash Image)

The skill resolves its own location — set $ART_SKILL once so every snippet works whether the skill is installed at ~/.claude/skills/art or baked into the image at /opt/agentbox/skills/art:

ART_SKILL="${ART_SKILL:-$( [ -d /opt/agentbox/skills/art ] && echo /opt/agentbox/skills/art || echo ~/.claude/skills/art )}"

bun run "$ART_SKILL/tools/generate-image.ts" \
  --model nano-banana-2 \
  --prompt "[PROMPT]" \
  --size 2K \
  --aspect-ratio 1:1 \
  --output /path/to/output.png

Alternative Model

Model Selection Guide

Nano Banana 2 (default):

• Best for: Most image generation tasks, fast iteration
• Strengths: Pro-level quality at Flash speed, reference images, web search grounding, 512px-4K range
• Use when: Default choice for all standard image generation
• API model: gemini-3.1-flash-image-preview

Nano Banana Pro:

• Best for: Professional asset production, maximum reasoning
• Strengths: Advanced reasoning, multi-turn refinement, style transfer
• Use when: Quality matters more than speed, complex compositional tasks
• API model: gemini-3-pro-image-preview

API keys in: ~/.claude/.env

• GOOGLE_API_KEY - Required for both models
• REMOVEBG_API_KEY - Background removal (optional)

Quick Preview Workflow (512px)

Use --size 512px for fast, cheap previews before committing to full resolution:

# 1. Preview at 512px (fast, low cost)
bun run "$ART_SKILL/tools/generate-image.ts" \
  --prompt "[PROMPT]" --size 512px --output /tmp/preview.png

# 2. Happy? Regenerate at full size
bun run "$ART_SKILL/tools/generate-image.ts" \
  --prompt "[PROMPT]" --size 2K --output /path/to/final.png

Thinking Flag (Nano Banana 2 only)

Add --thinking to give the model more reasoning time for complex compositions:

bun run "$ART_SKILL/tools/generate-image.ts" \
  --prompt "Complex multi-element scene..." \
  --thinking high --output /path/to/output.png

Only works with nano-banana-2. Thinking is always on (minimal by default). Use high when composition accuracy matters more than speed.

Common Diagram Pitfalls & Platform Limits

Generating an architecture/technical diagram, or an image for a specific social/publishing platform? Read references/output-constraints.md — it holds the diagram-pitfall table + pre-flight checklist and the per-platform dimension/size constraints with compression commands.

Site-Specific Style Override (CRITICAL)

Before generating images for a specific website, CHECK for style guides:

# Check project root for style documentation
ls HERO-IMAGE-TEMPLATE.md    # Site-specific image style
ls STYLE-GUIDE.md            # Design system

If found, the site's style guide OVERRIDES the default aesthetic.

Default aesthetic.md applies when: No site-specific style guide exists.

Consistent Dimensions for Related Images

When generating multiple images for the same page/section:

• Use identical resize dimensions for all images
• Document the size used: # All diagrams: 800x500
• Prevents distortion when displayed together

Nano Banana Prompting Guide

For detailed prompting techniques: nano-banana-guide.md Applies to both Nano Banana 2 and Nano Banana Pro — same prompt patterns, same API structure.

Includes:

• Core prompt formula: [Action] the [Subject] by [Specific Change]. The goal is [Desired Outcome].
• Action verb vocabulary (recolor, retouch, style, adjust, enhance, transform, add, remove, replace, blend)
• Mood and atmosphere vocabulary tables
• Color palette phrases
• Aspect ratio selection guide
• Iterative refinement workflow
• Brand integration template

Quick Decision Tree

What does user need?

├─ Unsure which approach? → VISUALIZE (analyzes & orchestrates)
├─ Flowchart/sequence/state diagram? → MERMAID
├─ Abstract metaphor for article? → Editorial (workflow.md)
├─ System/architecture with labels? → Technical Diagram
├─ Categories in grid? → Taxonomy
├─ Change over time? → Timeline
├─ 2x2 matrix or mental model? → Framework
├─ Side-by-side contrast? → Comparison
├─ Markup existing screenshot? → Annotated Screenshot
├─ Step-by-step process? → Recipe Card
├─ Sketchnote or visual meeting notes? → Sketchnote
├─ Quote as social visual? → Aphorism
├─ Idea territories as map? → Conceptual Map
├─ Single striking number? → Stat Card
├─ Multi-panel story? → Comic
└─ Edit/modify/transform existing image? → IMAGE EDITING

LinkedIn Slide / Carousel Production

To turn results into a stunning, faithful, LinkedIn-ready slide deck (Beamer 4:3 → MuPDF rasterise → Nano Banana Pro per-slide upcycle with a shared metaprompt → audit → composite PDF), see references/linkedin-slides.md. Reusable tooling: tools/slide-upcycle.sh + tools/slide-upcycle-metaprompt.txt. Key rule: diagrams-as-code (PGFPlots) give exact charts; the upcycle improves only the visuals — the --ref image is authoritative, every number/label is reproduced verbatim.

Diagram Upcycling (TikZ / Mermaid / matplotlib → publication quality)

Enhancing a rendered diagram for print/publication? See references/diagram-upcycling.md — render-to-PNG, upcycle (via generate-image.ts with nano-banana-pro, or the current google-genai SDK with gemini-3-pro-image-preview), ImageMagick fallback, the ≥2x text-hallucination rule, and LaTeX integration.

Assimilation Note

Source: Personal_AI_Infrastructure Art Skill Assimilated: 2026-01-04 Adapted: PAI's dark/neon aesthetic → warm/cream aesthetic

For complete visual styling rules, ALWAYS read: aesthetic.md