| name | prompt-to-design |
| description | Generate a polished PNG graphic from a text prompt and an aspect ratio. This is a CODE-BASED design generator (not a diffusion/photo model): it builds images from gradients, mesh fields, glow, grain, geometric shapes, and real typography. Use it whenever the user wants to create / generate / make an "image", poster, cover, wallpaper, banner, thumbnail, album art, story or reel cover, quote card, or any graphic from a description — optionally with words to render and a size or ratio (e.g. 1:1, 4:5, 9:16, 16:9, 2:3, 1:2, or WxH). It excels at designed, typographic, abstract, gradient, and geometric/Bauhaus/Swiss graphics and soft product-style visuals. It CANNOT produce photorealistic photos, real named people, brand logos, or copyrighted characters — for those, translate intent into a strong designed/stylized take instead. |
prompt-to-design
Turn a free-form prompt + an aspect ratio into a finished PNG, built entirely from
code. The house style is the level of the four pieces in reference/: posters,
typographic covers, mesh/gradient atmospheres, geometric compositions, soft
product abstractions. Frictionless like an image generator, but bounded to
designed graphics.
What it is (and is not)
Everything is drawn with math and type — no image model. That means:
- Great at: posters & album art, quote / story / reel / carousel covers,
wallpapers, gradient & mesh backgrounds, neon / synthwave, Bauhaus / Swiss
geometric art, minimalist logos-as-shapes, patterns, soft "product" orbs,
anything typographic. Crisp text, perfect alignment, editable, no AI artifacts.
- Cannot do: photorealism, a specific real person's face, detailed
representational illustration (a recognizable animal, a fantasy castle), brand
logos, or copyrighted characters.
Accept any prompt and always output a designed PNG. Silently route the prompt
to the rendering approach that fits (see Style routing). Only if a prompt truly
demands a photo (e.g. "photorealistic portrait of my friend") say so in one
sentence and deliver the strongest stylized/typographic interpretation anyway —
never refuse, never return a blob that ignores the prompt.
Inputs
- Prompt — free text. Extract: subject/scene, mood, any color cues, and
any literal words to render (quotes, titles, handles), plus implied style.
- Aspect ratio — a preset or
WxH. If unspecified, default 1:1.
The pipeline (follow in order)
- Read the prompt. Identify subject, mood, palette hints, literal text, ratio.
- Route to a style (one of: typographic, geometric, gradient-atmospheric,
soft-minimal, poster-scene) — see below. When ambiguous, pick the one that
best serves the words + mood.
- Choose a palette (3–6 colors, deliberate). See Palette.
- Plan the composition in your head first: the ONE focal element, the
hierarchy, where the negative space lives. Balance mass with space.
- Assign fonts by role (display / grotesque / geometric / serif / serif_book /
mono / pixel / techno). See Typography.
- Write a Python script that imports the engine and composes the piece
(pattern below). Prefer the library primitives; drop to numpy only for custom
effects.
- Render to a PNG in the outputs directory.
- VERIFY (mandatory gate — see Verify). Fix issues, re-render.
- Present the PNG. (Do not show the user the internal design reasoning or any
.md notes — PNG only.)
Optionally, thinking through a one-paragraph "design philosophy" for the piece
before coding measurably improves results — do it as internal reasoning, never as
a user-facing file.
Environment & how to run
Requires Python with Pillow and numpy (install with
pip install pillow numpy --break-system-packages if missing). Fonts are bundled
in fonts/ — no system fonts needed.
Write a script that puts this skill's lib/ on the path, composes, and saves:
import sys
sys.path.insert(0, "SKILL_DIR/lib")
from render import Design, NEON, NEON_CYAN, CREAM, INK, CHARCOAL, lerp_stops
d = Design("9:16")
d.save("/mnt/user-data/outputs/design.png", grain=6, saturation=1.1)
Replace SKILL_DIR with the absolute path to this skill. Then run with
python3 your_script.py and present the saved PNG.
Engine quick reference (render.Design)
Coordinates are fractions: x in [0,1] across width, y in [0,1] down height.
Sizes / radii / widths / font sizes are in final pixels; the engine
supersamples internally and downsamples for crisp edges.
Construct: Design(size="1:1", supersample=None, background=(255,255,255))
Backgrounds (set the whole canvas):
fill(color)
linear_gradient(stops, angle=90) — stops=[(pos,(r,g,b)),...]; 90=top→bottom, 0=left→right
radial_gradient(stops, center=(0.5,0.5), radius=0.9)
mesh_gradient(points) — points=[(fx,fy,(r,g,b),sigma_frac),...] smooth blended color fields
overlay_glow(center,color,radius,strength=0.4,mode="screen"|"add"|"blend") — soft glow onto current canvas
vignette(strength=0.4, center=(0.5,0.5), radius=0.75, power=1.6)
Shapes (optional glow={"color":(r,g,b)} for neon):
disk(cx,cy,r,color,glow=) · ellipse(cx,cy,rx,ry,color) · ring(cx,cy,r,width,color)
pie(cx,cy,r,a0,a1,color) — filled arc/sector; angles: 0=east, CW; top half=180→360, bottom=0→180
arc(cx,cy,r,a0,a1,width,color) — open stroke arc
rect(x0,y0,x1,y1,color,radius=0) — radius>0 for rounded/pills
line(x0,y0,x1,y1,width,color) · polygon([(x,y),...],color)
intersection(shapeA, shapeB, color) — knockout the overlap; shapes are
("disk",cx,cy,r), ("pie",cx,cy,r,a0,a1), ("rect",x0,y0,x1,y1)
Soft orb: gradient_sphere(cx,cy,r, colors, light=(-0.34,-0.34), shadow=True, specular=0.16, rim=0.05, glow=None) — colors = 2–3 stops blended diagonally.
Type:
write(x,y,text, role=, weight="regular"|"bold", italic=False, size=, color=, gradient=None, align="left"|"center"|"right", tracking=0, glow=None, shadow=None, stroke=0, stroke_color=None, return_width=False) — one line; y is the baseline.
Pass gradient=[(pos,(r,g,b)),...] for chrome/metal fills; glow={"color":...}
for neon; shadow=True or a dict for legibility on busy backgrounds.
text_block(x,y,text, ..., max_frac=0.84, line_height=1.16) — auto-wraps & stacks.
fit_size(text, target_frac, role=, weight=) → largest size that fits that width.
wrap(text, max_frac, ...) → list of lines. measure(text, ...) → (w,h) in final px.
Custom numpy: coords()→(xx,yy) · get_rgb() · set_rgb(arr) · composite_rgba(arr).
Finish: save(path, grain=0, chroma=0, scanlines=0, saturation=1.0, contrast=1.0, brightness=1.0) — post effects are applied after downsample (correct place for grain).
Typical: light designs grain=2; rich/dark designs grain=5–6, chroma=2; add
scanlines=0.03 only for CRT/retro looks.
Style routing
- typographic — prompt centers on words/a quote/a title, or asks for a "quote
card", "cover", "poster with text". Huge fitted type is the hero over a gradient
or mesh. Emphasize one word (color or serif italic). (ref: quote_card.py)
- geometric — "Bauhaus", "Swiss", "geometric", "shapes", "minimal poster",
abstract mark. Flat color fields, circles/arcs/lines on a grid, knockouts,
generous whitespace. (ref: geometric_bauhaus.py)
- gradient-atmospheric — "gradient", "mesh", "aurora", "synthwave", "vaporwave",
"neon", "dreamy", "abstract background", wallpaper. Mesh/linear fields + glow +
grain; add shapes/type as needed. (ref: synthwave_poster.py)
- soft-minimal — "soft", "Apple", "clean", "pastel", "premium", "carousel
slide", "product". Light airy ground, a soft
gradient_sphere or rounded card,
refined type, lots of space. (ref: soft_carousel.py)
- poster-scene — a simple scene ("sunset over mountains", "ocean horizon",
"desert dunes"). Render as flat/geometric layers: gradient sky, a
pie/disk
sun, polygon mountains, layered bands — a stylized designed landscape, not a
photo. Combine gradient + shapes.
Most prompts blend two (e.g. a synthwave poster is gradient-atmospheric +
typographic). Compose accordingly.
Palette
Pick 3–6 colors that carry a clear mood; restraint reads as premium.
- Warm/energetic: vermilion, coral, gold, magenta.
- Cool/calm: indigo, cobalt, teal, periwinkle.
- Neon: hot magenta + cyan on near-black.
- Soft/premium: cream/paper ground, muted pastels, one gentle accent.
- Bauhaus: cream + primary red/blue/yellow + ink (+ optional teal).
Use a light ground for airy/editorial; a dark or saturated ground for punchy/neon.
Ensure text contrast (see Verify). Grain/dither prevents banding in smooth fields.
Typography
- Match the face to the mood:
display/display_chunky for huge poster words;
grotesque/geometric for clean modern/Apple; serif/serif_display/serif_chic
for editorial/elegant; serif_book italic for a graceful accent word;
mono for kickers/handles/labels; pixel for retro-arcade; techno for sci-fi.
- Make hero type genuinely large — use
fit_size so the longest line fills the
column. Break a phrase into short lines for tall formats.
- Mix sparingly: one hero face + one label face; emphasize at most one word.
- On busy/gradient grounds, give type a soft
shadow or glow for legibility.
- Add a small letter-spaced kicker and/or a small footer to sell "designed piece"
— but keep them quiet.
Verify (mandatory before presenting)
After rendering, always check the result. Prefer viewing the PNG. If the image
viewer is unavailable or returns nothing, fall back to measuring pixels (this is
reliable and catches real bugs):
- Legibility: sample the background luminance behind the text vs the text
color — ensure strong contrast. Light text needs a dark-enough ground (and/or a
shadow/glow); dark text needs a light ground.
- Presence & placement: scan for the key colors (e.g. count pixels near the
accent/headline color) and print their bounding box to confirm elements landed
where intended and nothing was accidentally covered or knocked out.
- Composition: downsample to a ~44-wide grid and print a coarse "map"
(nearest-palette letters, or a luminance ramp) to see balance, coverage, and
whitespace at a glance.
- Banding: ensure grain/chroma is on for smooth gradients.
Example checks:
from PIL import Image; import numpy as np
a = np.array(Image.open(path).convert("RGB")).astype(int)
lum = 0.299*a[...,0]+0.587*a[...,1]+0.114*a[...,2]
band = lum[int(.3*a.shape[0]):int(.7*a.shape[0])]
print("bg lum behind text:", round(band[band<200].mean()))
mask = (((a-np.array(ACCENT))**2).sum(2) < 1500)
ys,xs = np.where(mask); print("accent bbox:", xs.min(),xs.max(),ys.min(),ys.max())
Common bugs to look for (all seen in practice): a knockout intersection that
swallows a whole shape because one shape sits fully inside the other; text/caption
placed on same-color area so it's invisible; a hero element too small; a gradient
that's so dark/muddy the colors don't read (lift it / boost saturation). Fix and
re-render until it's clean.
Aspect ratios
| Preset | Pixels | Typical use |
|---|
| 1:1 | 1500×1500 | square post (default) |
| 4:5 | 1080×1350 | IG portrait |
| 9:16 | 1080×1920 | story / reel / phone wallpaper |
| 16:9 | 1920×1080 | slide / desktop / YouTube |
| 2:3 | 1200×1800 | poster portrait |
| 3:2 | 1800×1200 | poster landscape |
| 1:2 | 1080×2160 | tall poster |
| 3:4, 4:3, 5:4, 2:1 | — | also available |
Or pass WxH (e.g. "1600x1000"), clamped to a 2560 long edge for speed.
Guardrails
No photorealistic depictions of real, named people; no reproductions of brand
logos or trademarks; no copyrighted characters or existing artworks; nothing
harmful. These are designed original graphics. When a prompt asks for one of
these, deliver an abstract/typographic/stylized interpretation instead and say so
briefly.
Reference examples
Study reference/ — each is a runnable, gold-standard build for a prompt type:
synthwave_poster.py — gradient sky, striped sun, neon grid, chrome title (1:2)
geometric_bauhaus.py — flat fields, arcs, knockouts, dot grid (1:1)
quote_card.py — mesh gradient + huge type + italic/gold accents (9:16)
soft_carousel.py — light ground, soft orb, refined type, page dots (1:1)
When a new prompt resembles one, start from its structure and adapt palette,
composition, copy, and ratio.