Calling an LLM from Swift is easy. Keeping the workflow around that call understandable and maintainable is the hard part.
Shikisha is my attempt to fix that. It gives small chains, RAG pipelines, tool-using agents, and app UIs the same basic vocabulary, so prompts, structured output, retries, streaming, tools, and tests stop being a per-project pile of glue code.
It is still early, currently around 0.1.0, but it has reached the point where I reach for it in real projects instead of sketching examples around it. JPResume is the first place it earns its keep.
The interesting part has not been the model calls. It is what happens to the rest of the system once those calls become one composable step instead of the whole thing.