For those of you who use LLMs to give you touchdesigner tips, what is your workflow?
For a long while I’ve used various chatbots, whether it was ChatGPT or Claude Opus 4.5 inside of VScode/Windfsuf, to answer design questions about Touchdesigner. Ex. “how do I apply easing to a 0-1 ramp signal using CHOPs”.
And these chatbots tend to hallucinate badly with Touchdesigner in particular. 3/4 times they recommend using operators that simply don’t exist in TD (often the names are coming from Houdini) and I’ve been struggling to get value out of them.
Have people figured out a better solution? Maybe some clever (or simple?) way to encapsulate a lot of TD-specific context – such as copies of all the docs – to make common models smarter when it comes to TD specifics?
There is a lot of TD information online. But I don’t know the right way to turn this information into actionable context, to improve the quality of TD-related design suggestions from AI. I don’t have this problem when asking about common software, like ansible / python / shell, or even other creative software like Blender. There is something about TD that really confuses these systems (well, confuses me too, but that’s why I’m hoping to get some design tips from AI for it haha).
Because TD is not a typical text based programming framework, most LLMs have trouble contextualising it from their own training data, so you need to give it more to work with
Uploading various reference documentation and giving it specific URLs (wikis, docs, forums, etc…) seems to yield better results, never perfect, but far better than not providing them.
I recently switched to Gemini from ChatGPT because it’s general guidance on all things technical has been much more useful.
ChatGPT would tend to invent TD operators about half the time whereas the Gemini thinking model seems to rarely do that. It’s certainly not perfect but it’s the best I’ve tried so far.
Not used it for that tbh, but I am regularly impressed with how useful Notebook LM can be.
It only uses sources that you feed it, but that can be Youtube Urls and it could be all official documentation etc. Maybe it’s worth a try?
I did a custom ai bot where you can pass the whole TD offline documentation (RAG) and a some very detailed agents prompts. The result is ok if you want to learn and discuss python, or what a node does but certainly not to build a patch. Or suggest how to wire nodes.
Local model needs to be huge, i got good results from 15-20gb models on my 4090. So to save you time:
