I’ve also found that LLMs are great for getting past developer’s block. “What next?” can overcome inertia quite easily.

Yup, I don't use models to write serious code. But if my brain is totally blocked, perhaps after half day of meetings, I do sometimes take a smaller local model, explain a simple task to it, and let it try to implement to my specs. At this point I'm usually annoyed and engaged enough to get back in the zone.

If a human is available, 30 minutes of pairing works even better. There's just something about breaking tasks down and getting even simple feedback that makes it a good jumpstart.

That and I find it’s usually easier to start with a shitty first attempt that kind of works that you’re then refactoring