You can almost certainly use GPT-4 or Claude 3.5 to help design a spec. But there’s more to spec design than just whipping spells and talents in a vacuum. The model needs to grok spec design theory, especially applied to the game you’re designing for, and you need to help it grok that. So that it can build up synergy between abilities and talents. Procs, buff windows, those kinds of things.
I suspect a better prompting strategy would revolve around first feeding the context with spec guides for Havoc and Vengeance. Icy-Veins would do nicely, because these go in depth explaining the why of talent choices and rotations. Then feed it the same for one or two healer specs, so it understands what makes a healer tick.
Then verify if it understood synergies between abilities, talents, procs, buffs, and how they drive rotations and choices. Ask it to give specific examples. You do this to verify that it understands what makes spec design work.
Only then, ask it to design a Demon Hunter healing spec of its own. Ask it to think step by step. For example, the theme of the healer should be “Mobility” first, “Convert damage to healing” second: being able to heal during movement intensive combat, and being able to convert a fraction of damage dealt by the Demon Hunter or her teammates into healing. First, it should leverage a few existing damage abilities that are already common between Vengeance and Havoc; explaining why it retains these abilities. Second, it should create healing abilities; it should explain why it designed the mechanics the way it did, and how the abilities complement each other. While exploiting existing abilities and designing new ones, it should try to create a rotation and synergy between the abilities, enhanced by talents. All the while maintaining the themes and niches of mobility and damage converted into healing.
Those steps should give you a decent first draft for ideas. But modern spec trees are big. Also, you might stumble upon weird pitfalls: for example, maybe the model doesn’t understand that movement usually interrupts casting. So perhaps you’ll need to tweak your prompts to clarify that, or ask it to clarify after a first draft so it understands the mistakes it made. You’d need to iterate and refine a lot from that starting point to get anything production ready.
Credentials: accustomed to use LLMs in an official capacity for work.
