SVG

An SVG is code, so it comes from a delegate text tier rather than an image model: it costs plan allowance instead of the small daily image quota, and the result is reviewable, diffable and editable. svg.sh then does the part a model cannot be trusted with — stripping anything executable, fixing the size, and making the loop last exactly as long as you asked.

Quick start

svg.sh make "a ringing notification bell" bell.svg       # 32px, transparent, CSS, 1s, forever
svg.sh still "a bell" bell.svg                           # one static image, no animation
svg.sh make "a walking robot" walk.svg --frames 6        # frame-by-frame: 6 drawings, timed here
svg.sh make "a loading spinner" spin.svg --size 64 --duration 1.5
svg.sh build out.svg reply.txt --still                   # normalize SVG you already have (or -)
svg.sh probe out.svg                                     # what the file contains; exits 1 if unsafe
svg.sh preview out.svg out.png --width 512               # rasterize frame 0 so you can look at it
svg.sh check                                             # dependencies

Defaults: 32x32, transparent, CSS-animated, 1.0s cycle, loops forever. Sizes are restricted to 16 / 32 / 64 / 128 and duration to 3s.

Three ways to animate — pick by where the file is going

--still (or the still verb) drops animation entirely. All three animated modes are normalized to exactly --duration per loop.

What happens to the model's answer

Never trusted as-is. svg.sh extracts the SVG from the prose or fences around it, refuses a DOCTYPE outright, and strips <script>, <foreignObject>, every on* handler, javascript: URLs, external and data: hrefs, and CSS @import — then normalizes size and viewBox, applies or removes the background, scales every animation time so one loop is exactly --duration (preserving each element's phase rather than flattening the stagger), forces the iteration count so the loop actually loops, and adds a prefers-reduced-motion rule. A reply with no animation and no --still or --frames is rejected rather than quietly shipped.

In --frames mode the model only draws: it is asked for N sibling <g> groups and forbidden to animate them, and the script writes the step-end keyframes and the per-frame negative delays itself. Models get that arithmetic wrong; they do not get "draw the same robot mid-stride" wrong.

Rejections are re-asked, once by default, with the error handed back to the model and the tier escalated (code → hard). Retries cost plan allowance, not your context.

Working with it

Look at the result before reporting it done. svg.sh probe tells you what is in the file; it does not tell you whether the drawing is any good. svg.sh preview out.svg out.png rasterizes frame 0 so you can actually read the image — do that. The animation itself only plays in a browser.

When a result is close but wrong, --keep-raw keeps the model's reply so you can edit it and re-run svg.sh build instead of regenerating. Editing the SVG by hand is fine too — it is just code — but run it back through svg.sh build afterwards so size, timing and safety are re-enforced.

For a subject that does not animate naturally (a static logo, a flat glyph), use --still rather than accepting a token wobble.

Detail on the sanitizer, the retiming maths and every failure mode is in REFERENCE.md. Needs python3 only (stdlib — no Pillow, no ImageMagick), plus delegate.sh on PATH for generation.