프롬프트 체인

A single mega-prompt asks the model to do everything at once: read the source, extract facts, reason about them, follow style rules and format the answer. Each extra demand competes for attention. When the output is wrong, you cannot tell which part failed. Prompt chaining splits the work into stages, and each stage has one clear job. A typical chain has three steps: 1. Extract the relevant quotes from a document. 2. Answer a question using only those quotes. 3. Check the answer against the quotes and rewrite it for the audience. Ordinary code can check outputs between steps. It can parse JSON, confirm that a required field exists, or stop the chain if a classifier returns low confidence. Anthropic's December 2024 essay "Building Effective Agents" calls these checks "gates." It describes prompt chaining as a workflow that trades extra latency for higher accuracy, because each call becomes an easier task. Chains and agents differ in who controls the path. In a chain, the developer hard-codes the sequence. In an agent, the model decides at run time which tool to call next and when it is done. Chains are cheaper to test, easier to explain and fail in predictable places. Agents suit open-ended problems where the steps cannot be known in advance. Many production systems are chains rather than agents, sometimes with branches (routing) or steps that run in parallel. Three misconceptions are common: - Chaining always costs more. In fact, small focused prompts can use cheaper models for the easy steps. - More steps are always better. Every handoff can lose information or compound errors. - Chaining needs a framework. Libraries such as LangChain offer helpers, but a chain is just a few function calls passing strings or JSON.