Orchestration - chaining AI calls
Complex AI systems rarely use one giant prompt for everything - they break a task into smaller steps, each handled by a separate, narrower AI call, with the output of one step feeding into the next. Decomposition is the core orchestrator skill.
Chaining smaller, verifiable jobs (classify → extract → draft, for example) is more reliable than one massive prompt trying to do everything at once, because each step is easier to check, debug, and improve independently. Frameworks like LangChain, LangGraph, and CrewAI exist specifically to build and manage these chains in code. This is the actual mental model behind everything you already build - Kasper Trips, ProvaCV, your Claude skills library - even when it doesn't look like "AI agents" on the surface.
Take one of your own multi-step Claude skills (e.g. cv-tailoring) and map it explicitly as a chain: what's step 1's exact input/output, what feeds into step 2, etc. You'll likely find you've been orchestrating without naming it as such.
Why break a task into a chain of smaller AI calls instead of one giant prompt?