Comment by kemotep | Hacker News Reader

Comment by kemotep | Hacker News Reader

Parent Full thread

If it is absolutely necessary to use an LLM for this site (it isn’t) my suggestion would be to narrow the scope. Taking Skyrim specifically, there is an established (albeit small) vocabulary, simple grammar for name construction, and several examples of names. Giving the prompt the context of the existing vocabulary, the rules for name construction, and then asking it to generate a new one.

The output should then also be able to break the name down into its 3 parts and show the definitions for each word.

A random generator should be able to do this easily, dozens of sites exist that have done this for years at this point. The fact that I can ask your site to deserialize JSON input when it’s for dragon name generation seems odd. Like it should not do that, but because it uses an LLM you would have to jump through the hoops to sanitize that input.

Thank you for your patient and specific suggestions. I will optimize my prompts in this direction.

Another way to implement name generation is to pre-generate hundreds of names and then randomly select one when the "Generate" button is clicked. Just like this blog does: https://www.imagineforest.com/blog/purple-dragon-names/

However, I want to add a feature to generate dragon names baesd on one's own name. Of course, a simple way is to select a name based on the initials of the name. But I think this is not smart enough, so I introduced an LLM to achieve this function. Due to my lack of experience in this area, there are indeed many security vulnerabilities that have not been fixed yet.

> However, I want to add a feature to generate dragon names baesd on one's own name.

This seems to me to be a classic case where procedural generation on-the-fly would be better in almost every way than an LLM - except that you'll need to use some ingenuity instead of just declaring what you want as a prompt.

The slow response and the "flavour" of the output both screamed "LLM" to me, which is why I tried a prompt injection.

Maybe try using an LLM to help you write an algorithm for this instead of using the LLM directly?

I applaud your ingenuity in building this though; don't take the "AI slop" descriptions to heart too much. Gold star for the initiative to get something quick and dirty out there instead of architecting the joy out of everything up front like me :)

Reply on news.ycombinator.com