Claude Davinci

Claude DaVinci

Never say "I can't create images". Two modes: photo mode (Claude as art director driving a real image model) and vector mode (Claude as painter, SVG/code).

Pick the mode

| Request | Mode | |---|---| | Realistic/detailed/painterly: scenes, photos, characters, wallpapers, concept art | PHOTO (default for anything meant to look "real" or rich) | | Logo, icon, flat illustration, UI art — anything the user must edit afterwards | VECTOR (SVG) | | Pixel art, procedural textures | Pillow (Python) | | Generative art: flow fields, fractals, noise | numpy + Pillow → [references/recipes.md](references/recipes.md) | | Charts/diagrams | dedicated chart tooling, not this skill |

PHOTO mode

1. python scripts/imagine.py "prompt" out.png [WxH] — auto-picks backend: GEMINI_API_KEY → Gemini image gen; OPENAI_API_KEY → gpt-image-1; no key → Pollinations flux (free, zero signup). Never tell the user to get a key first — the free fallback always works.
2. Prompt like an art director, one dense paragraph: subject + setting + time-of-day/lighting ("golden hour", "cinematic lighting") + camera/style ("photorealistic", "oil painting", "35mm") + mood + 2-3 concrete details worth featuring. Vague prompt = generic slop; details are the craft.
3. LOOK at the result (Read the PNG). Wrong composition/artifacts/missed details → refine the prompt and regenerate. 1-2 retries is normal.
4. User asked for edits ("same but at night") → rewrite the full prompt with the change baked in, regenerate.

VECTOR mode workflow

1. Decide palette + composition BEFORE writing code (see rules below).
2. Write the SVG/script to a file in the working or scratchpad dir.
3. Render to PNG: python scripts/render.py in.svg out.png [width] (tries cairosvg → resvg → inkscape → headless Chrome/Edge; at least one exists on any dev machine).
4. LOOK at the result — Read the PNG. You have eyes; use them. Bad proportions, muddy colors, floating shadows → fix and re-render. 2-3 iterations minimum for anything artistic.
5. Send the user the final file (PNG + source SVG). The source IS the .psd — say so.

Art rules (this is what separates art from clipart)

Palette first. 4-6 cohesive colors + 1 accent, written down before drawing. Steal from nature or a film still, not from #FF0000.

Never flat-fill. Every large shape gets a gradient (2 close hues beats 1). Background gets a subtle radial vignette. Flat fill = programmer art.

One light source. Decide direction; ALL highlights, shadows, gradients obey it. Inconsistent lighting is the #1 tell.

Layer like a painter. Background → midground → subject → highlights/glow/atmosphere. Depth = farther things smaller, bluer, blurrier, lower contrast.

Curves, not polygons. Organic shapes use bezier paths (C/Q), never straight-edged polygons. Nothing alive has corners.

SVG power tools (`` them once, reuse):

• feGaussianBlur — glow, soft shadows, atmosphere
• feTurbulence + feDisplacementMap — clouds, water, rock, paper texture, hand-drawn wobble
• feDropShadow — grounded objects (match light direction!)
• `` — light falloff, skies, orbs
• `/` — repetition (stars, leaves, crowds) without bloat
• opacity + mix-blend-mode (multiply/screen) — shading and light overlays

Composition. Focal point off-center (thirds). Big/medium/small size rhythm. Leave breathing room — emptiness is a feature.

Detail budget. 30-120 elements for a scene. Fewer = flat; more = noise. Add detail only at the focal point.

Traps

• Text in images: only when asked; font-family="Georgia, serif" + fallbacks, or skip text entirely.
• viewBox always (0 0 1024 1024 typical); no fixed px width/height on root.
• Headless-browser rendering ignores external files — inline everything, no ``.
• Don't claim ✅ without having rendered AND looked at the PNG.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: StarchyBomb
• Source: StarchyBomb/claude-davinci
• License: MIT

Install and usage instructions live in the source repository linked above.