Decomposition is what you do BEFORE writing the prompt, then encode IN the prompt. You look at the task, identify its N sub-problems, and write instructions that address each one in a specific order.
Distinct from prompt chaining (which splits across API calls) and chain-of-thought (which asks the model to reason step by step). Decomposition front-loads the structure: "First X, then Y, then combine into Z." The model doesn't have to discover the decomposition; you hand it over.
Example Prompt
Evaluate this customer feedback.
Address these four sub-questions in order:
1. What is the customer's specific complaint?
2. Is this a product issue, a service issue, or a policy issue?
3. What is the minimum action needed to resolve it?
4. What is the optimal action to delight the customer beyond resolution?
Produce a one-paragraph summary only after you've worked through 1-4.When to use it
- Complex tasks where the model has been picking the wrong decomposition
- You know the sub-problems better than the model
- You want the output to cover specific dimensions every time
When NOT to use it
- Simple tasks where the decomposition is obvious
- You don't actually know the sub-problem structure (let the model try first)
- The "decomposition" is really a laundry list of "also do these 10 things"