Anyone can now build an LLM demo with a single API call. Getting that same feature to survive real users, real data, and a real bill is a different job. That job comes down to five skills that show up in almost every production system I've worked on or reviewed: retrieval, routing, guardrails, evals, and agent loops.

Each one answers a question a frontier model alone can't handle. Where does my company's data come from? Why is the bill this high? What happens when someone sends a hostile prompt? Did my last change make things better or worse? How do I get a model to do multi-step work without falling over? Skill #1: RAG, grounding a model in your own data A base model doesn't know your company's documents, and fine-tuning on them every time they change is slow and expensive. Retrieval-augmented generation (RAG) fetches the relevant documents at query time and passes them to the model as context.

It's still the most common LLM pattern in production, and it's usually the first thing a curriculum teaches. The interesting work is in the retrieval. Whether you return the right chunk depends on how you split documents (chunking), whether you combine keyword and semantic search (hybrid retrieval), and whether you reorder results before they reach the model (reranking).

Skill #2: Model routing: paying for the model the task needs Sending every request to a frontier model is the fastest way to a bill nobody can explain. Routing sends easy, high-volume requests to a small or local model and saves the expensive model for the hard ones. Adding a cache for repeated prompts is reported to cut 40% to 70% off production inference bills, which is why routing went from an advanced trick to a baseline expectation.

The skill is scoring a request and picking a tier.