just one more prompt bro

But what if you actually did that? What if you actually just kept prompting?

Well you won't be able to write any web browsers or TLS implementations, but you can get much further than you think, as long as there is a sense of direction. Two examples from M3 Svelte:

Making a nice looking ripple

Material 3 has an interesting ripple. It's not flat like previous versions or shiny like plastic - it's closer to how water ripples when you touch it (see this clip) thanks to its dithering. I wanted to implement this with SVG, but I had no idea where to start. So in November 2024, I asked AI (some GPT I think) "make a ripple in svg that goes from complete opacity with no noise for a bit to a noisy fade to complete transparency".

Of course its first try didn't work. Same goes for its second, and its third, and so on. But I kept telling it how its render failed, and kept telling it to try something else. It had many options to try - SVG is a large spec.

That's how I found out about <feDisplacementMap>. It shifts pixels around, and combined with <feTurbulence>, you get this great ripple effect:

Really, used in M3 Svelte

Making automatically transitioning paths

One of SVG's limitations is that 99.99% of the time, you can't morph one path to another. Most folks pull in something like flubber, which doesn't use the browser's native path interpolation and weighs 19kB (post-compression!)

Another approach is precalculation, doing something like "generate the path at 0.1% and 99.9%", but then you have to run this for every single pair of paths you want to animate between.

So in November 2025, I uploaded the Material 3 shape library and my SVGs from there to Claude (since I'm on Pro, it has a computer), and asked it to make something better. You can view the chat here. Of course, this wasn't a single shot - I first came up with the idea of converting every shape to a polyline-like path with a fixed number and constant arrangement of points to allow universal morphing, then led it to represent every shape in this format, then ran a smaller scale trial comparing this to Flubber and debugged a few bugs, then ran a larger scale trial on all shapes, guiding it to improve accuracy while keeping animation working and file size low.

Similarly, I got this working!