Lesson 03 · Workflows · 11 min
Routing
A single prompt that handles every kind of input has to hedge. A shipping question, a refund request and an angry escalation want different facts, different tone and different constraints, and a prompt that covers all three is worse at each.
Routing separates the concerns. A cheap first call classifies the input into a small closed set; a specialised prompt then handles each class. The specialist can be narrow and specific — the shipping prompt knows the delivery times and is forbidden from inventing a date — because it will only ever see shipping questions.
The classification is the weak point. Keep the set small and closed, and constrain the classifier's output to exactly those words, because a free-text label cannot be matched reliably. Always provide a default route for when the classifier produces something unexpected — in the lab it falls through to the escalation specialist, which is the safe direction to fail.
This is the graph course's routed graph, seen from the pattern side. There the routing is drawn as nodes and edges; here it is a classifier call and a lookup table. Same idea, and the graph version is what you reach for when the routes need to be inspectable or resumable.
Routing also lets you route by cost. An easy class can go to a small fast model and a hard class to a large one, which is often the single biggest saving available in a system that handles varied traffic.
Checkpoint
Find a task where the specialist gives a more accurate answer than the single call, and say what it knew that the generalist did not.