Prompt chaining and routing

Prompt chaining

Break a task into steps. Each step's output feeds the next. Put a gate (a plain code check) between steps so a bad intermediate result stops the run early instead of poisoning everything after it.

outs = chain(llm, [("summarise", "SUMMARISE\n{input}", None),
                   ("extract",   "EXTRACT\n{input}",   lambda s: "no order id" if "unknown" in s else None)], ticket)

• Use when the task splits cleanly into fixed stages and each stage is easier than the whole.
• Cost: one call per step, and the latency adds up.
• Fails by passing along an early mistake. Gates, and validating each output (Level 2, module 3), are the fix. In the tests a failed gate stops the chain after one call.

Routing

Classify the input, then send it to a specialised prompt (or to a cheaper or stronger model).

label, reply = route(llm, text, "CLASSIFY\n{input}", handlers, default="other")

• Use when inputs fall into distinct kinds that need different instructions or tools.
• Cost: always two calls (classifier plus handler), but each prompt is shorter and focused.
• Fails by misclassification. Always give a default route, and measure the classifier on its own (module 5). The router tidies the classifier's reply (case, trailing full stop, quotes) before looking it up, because models often answer Billing. when you asked for billing.