Prompt chaining is a pattern for building LLM applications in which a task is broken into a sequence of subtasks, each handled by its own prompt or model call, with each step working on what the previous one produced and the results integrated at the end. Anthropic, in Building effective agents, presents it as one of the compositional workflows it has seen in production agentic systems, and adds that programmatic checks — a “gate” — can be placed on any intermediate step to keep the process on track. A Chinese-language overview of agent construction by the developer phodal describes the same idea as a way to decompose complex problems systematically, and treats it as a companion to Prompt Routing: once requests can be routed, a complex one can be split into a chain.
Usage
The two accounts stress different benefits. Anthropic frames the pattern as a trade: accepting higher latency in exchange for accuracy, by making each LLM call an easier task. Its examples are generating marketing copy and then translating it into another language, and writing a document outline, checking that it meets certain criteria, and then writing the document from the outline.
phodal’s overview frames it as modularity: each subtask concentrates on one stage, any subtask can be rewritten or have its prompt replaced as needed, and later prompts can be adjusted according to what an earlier stage produced. Its example is turning a product manager’s idea into requirements through four stages — collecting ideas and initial requirements, sorting out the requirement logic and feature priorities, drafting a preliminary requirements document or task list, and finalizing the requirements into a formal document — with each stage able to be handled by a different prompt or sub-agent, such as a search-capable agent for the first.
When It Applies
Both sources place it where the steps are known in advance. Anthropic describes it as ideal when a task can be easily and cleanly decomposed into fixed subtasks; phodal describes it as suited to processes with a fixed flow in which some steps can be skipped. It assumes that the decomposition is known before the task starts. Anthropic’s stated cost is latency, since the task is spread across several sequential calls.
The pattern’s standing rests on practitioner accounts rather than measurement: Anthropic presents it as a pattern it observed across teams building agents, and phodal as teaching material drawn from training developers to build agents.
Related Terms
- Prompt Routing — the companion pattern that sends an input to a specialized path
- Augmented LLM — the building block each call in a chain is assumed to be
- Prompt engineering — the discipline phodal’s overview places it under