Magic Music Visuals

Hi there!

i am excited about your new ISF AI prompt and wanted to ask about some insights how to handle it. I had some error messages that confused me, I had some very basic results after prompting that I couldn't use and sudden few real great results.

first, the error messages, I am not sure which prompts I put into, I can track more accurate in the future, but I made 2 screenshots of the error messages attached. will those failed prompts still be charged on the AI credits?

this, or similar, was a prompt and I had some very different results with the different models and alternating versions of the prompt:

"You're on a dance floor, looking down at the floor and seeing the lights from the spotlights hanging above you, shining down on the floor, flashing and blinking in time with the music."

9 out of 10 tries with different models and prompt versions resulted in very basic designs, nothing worth showing. But one time this was the result, really great. I worked with it and added some settings and controls and additional features, but the concept of the grid at the bottom and the alternating flashing patterns was completely made up by the AI. using gemini-3.5-flash.

the result: https://www.youtube.com/watch?v=fAcaZgfXqMc
project and AI file attached. (sure you can see the prompts in the fs file in german)

I tried out more prompts for different visuals but I didn't get anywhere near that great result at all.

can you give me some advise how to use the prompt, what can I ask it and what shouldn't I? to avoid errors and to get good results like this one? should the prompt be rather short or long? should It be more detailed or less detailed? should I ask it from my personal view how it should look or should I describe more technical? I had the feeling that sometimes somehow a less detailed prompt in the AI node leaves more freedom in design. From other AI use cases I know though that the opposite should be the case.

cheers

This is a good example to work from. You're describing a POV - better to specify a shader itself. I would suggest thinking about the resultant 2d output more than an implied 3d environment, though I understand why you'd want to specify feels.

You have redundancy and ambiguity in your prompt (you say floor 3 times) which doesn't help focus the AI model. Swap narrative adjectives for precise lighting vocabulary. For example, your "Flashing and blinking" is ambiguous. It could be a hard on/off strobe and a continuous brightness pulse — be more explicit. Same idea for the beam itself: instead of "spotlights shining down," specify a *gobo* if you want a patterned/cut-out light shape, and a *penumbra* width (or falloff exponent) if you want to control how soft the edge is. These have a direct, unambiguous mapping to shader math — a gobo is a mask/stencil, a penumbra is a smoothstep falloff — so you're not leaving the interpretation up to chance the way you would with "soft" or "glowing."

ISF's whole advantage over a plain GLSL shader is exposed, linkable parameters — but the model only adds params it has a specific reason to add or if you ask explicitly. If you want live control over count, radius, penumbra, and palette, ask for those by name in the prompt. Also, think about what's available in Magic - for example you could specify "add intensity parameter with options for strobe, pulse, ramp" but for a simpler option with more flexibility you could just control 'intensity' with Magic modifiers.

example prompt wrote:Draw 2-10 light pools from gobo-shaped spots, each with a distinct penumbra falloff from bright center to fully transparent edge. If no input image is supplied default to a circle. Fixed positions, no camera movement. Assign each light pool an independent color from a warm/cool palette. Expose parameters: Count, Pool Radius, Penumbra Width, Palette (warm/cool blend), and an Intensity float I'll link to audio myself.

If you get an error or a poor result you don't get a refund. You can try to fix it yourself or submit it again with the errors to try to fix it. The devs have some control over these error cases but it takes time to find them and it's an ongoing process.

Sorry for these errors. We're always trying to make our context better by handling these kinds of things, but AI is not perfect.

Although Gemini 3.5 Flash is the least expensive and reasonably good, it's not the best and probably results in the most errors. I think Claude Sonnet (the default one) is the best balance of price, speed, and accuracy.

A good strategy might be to use the better, more expensive models (like Opus) for creating something completely new, and then the cheaper ones for further editing/refinement. I also generally start with a simple prompt and then add small changes as I go so that I can always undo.

I like your shader by the way -- it's a really nice one.

Any other questions, please let us know.