Prompt Chaining and Decomposition

One job per prompt

A single giant prompt that extracts, analyses, decides and writes is hard to debug. Prompt chaining splits it: step 1 extracts facts into JSON, step 2 checks them against rules, step 3 drafts the reply from the verified data. Benefits: each prompt is simpler and testable, you can use a cheaper model for easy steps, you can insert code checks between steps, and failures are easy to locate. Costs: more calls, more latency, and errors can propagate, so validate between steps. Routing (a first call that classifies the request, then sends it to the right specialised prompt) is a common chain.

A three-step chain

Each step has one job; code can validate between steps. Illustrative; not run here.

Step 1 EXTRACT   prompt: "Extract order_id, issue, requested_action as JSON."   -> validate JSON
Step 2 DECIDE    code/prompt: apply refund policy to the extracted fields    -> decision + reason
Step 3 WRITE     prompt: "Write a 3-sentence reply using ONLY these facts: ..." -> final text
Routing first: a cheap classifier sends billing / technical / other to different prompts.

Validate between steps

Check the JSON from step 1 before step 2 uses it, so one bad extraction does not poison every later step.

Quick check: What is a benefit of prompt chaining?

• It removes all errors
• Each step is simpler to test and debug
• It always costs less
• It removes the need for validation