Scripting all the things
For AI in genomics as it currently stands, there's an aspect I haven't seen mentioned that has been on my mind. In short, it's easier to provide more complex, powerful, and complete functionality to users without them being overwhelmed now that they can use an LLM. One way to expose more functionality is to provide a scripting interface to a command-line tool or low-level library.
Most often, knobs have been exposed in command-line tools as arguments, and often, not at all (think of all of the constants that are not exposed to the user). But now, given how good LLMs are at generating code, it's even more powerful to expose a scripting interface. Then, instead of 200 command-line arguments, it's possible to offer even finer control by embedding a scripting interface. Previously, this would have required more from users; now it requires that the scripting API be documented. We can then drop the docs into a custom GPT or a SKILL.md and the user can use the full power by writing plain-text prompts.
This can, of course, work for command-line tools without a scripting interface, but scripting gives LLMs a richer surface to work with. Likewise, Claude or Codex or Pi can help a user to modify low-level code, but if there's already a scripting interface, like there is for vcfanno or slivar, then it's quite powerful and reduces the maintenance burden. For example, a user can ask for a slivar expression that flags rare compound heterozygotes instead of learning every available field and helper function up front.
I've been doing this for some time. Get in touch if you have a relevant project.