decomposed prompting

**Decomposed prompting** is a prompt engineering technique that breaks a **complex task** into multiple **modular sub-tasks**, each handled by a specialized prompt or even a different model. Rather than asking an LLM to solve everything in one shot, you design a pipeline of simpler, focused steps. **How It Works** - **Task Decomposition**: Analyze the complex task and identify independent sub-problems. For example, answering "What is the market cap of the company that manufactures A17 chips?" requires: (1) identify the manufacturer → Apple, (2) look up Apple's market cap. - **Sub-Task Handlers**: Each sub-task gets its own optimized prompt, tool call, or specialized model invocation. - **Orchestration**: A controller (another LLM call or code logic) routes information between sub-tasks and assembles the final answer. **Key Benefits** - **Accuracy**: Simpler sub-tasks are individually easier for the model to get right, reducing compound error rates. - **Modularity**: Sub-task prompts can be **independently tested, debugged, and improved** without affecting others. - **Tool Integration**: Natural integration points for external tools — one sub-task might call a calculator, another might search a database. - **Transparency**: The reasoning chain is explicit and auditable, unlike monolithic prompts where reasoning is opaque. **Comparison with Other Techniques** - **Chain-of-Thought (CoT)**: Asks the model to reason step-by-step in a single prompt. Less modular. - **Least-to-Most Prompting**: Progressively solves sub-problems from simplest to hardest. More structured than CoT but less modular than full decomposition. - **Decomposed Prompting**: Each sub-task can use a **different strategy** — some might use CoT, others might call tools, others might use few-shot examples. **Real-World Applications** Used in complex **agentic workflows**, **multi-hop question answering**, **code generation** (plan → implement → test), and any scenario where a single prompt can't reliably handle the full task complexity.