Prompt routing is a pattern for building LLM applications in which an input is analysed first and then directed to whichever downstream path suits it best, rather than one prompt trying to handle every kind of request. Anthropic, which calls the pattern simply “routing” in Building effective agents, describes it as classifying an input and directing it to a specialized follow-up task. A Chinese-language overview of agent construction by the developer phodal defines prompt routing as an engineering pattern for multi-task, multi-agent or complex AI flows that splits tasks, analyses the input and assigns it to the most suitable model or sub-task prompt; its core idea, on that account, is to decide dynamically — from the input and its context — which processing path, prompt, tool or sub-agent to use, so that execution becomes conditional rather than linear.
Usage
Anthropic’s case for the pattern is separation of concerns: routing lets each path have a more specialized prompt, whereas without it, optimizing for one kind of input can hurt performance on others. Its examples are sending different kinds of customer-service queries — general questions, refund requests, technical support — into different downstream processes, prompts and tools, and sending easy or common questions to a smaller, cheaper model while harder or unusual ones go to a more capable one.
phodal’s example is a question-answering system: questions unrelated to the system are told they are unsupported, basic knowledge questions go to document retrieval and a QA model, and complex analytical questions go to a data-analysis tool followed by a summary. The overview names LangChain’s RouterChain, and routing by semantic similarity, as framework support for the pattern, and presents routing as what makes Prompt Chaining possible for complex problems.
When It Applies
Anthropic gives two conditions: the task has distinct categories that are better handled separately, and the classification itself can be done accurately, whether by an LLM or by a more traditional classification model or algorithm. Accurate classification is thus a precondition for the pattern, not a by-product of it. phodal’s account adds that routing lets a system pick the most suitable handling for each type of question while keeping it modular and extensible.
Both descriptions are practitioner accounts: Anthropic’s is drawn from patterns it reports seeing across teams building agents, and phodal’s from training material for developers learning to build agents. Neither reports a measured comparison with a single, unrouted prompt.
Related Terms
- Prompt Chaining — the companion pattern for decomposing a task into fixed steps
- Embedding-guided tool routing — a related technique that routes to tools rather than prompts
- Augmented LLM — the building block each routed call is assumed to be