One prompt doesn’t need to do everything.
Prompt chaining splits a task across separate LLM calls. Each call has a focused responsibility, and an intermediate result is consumed by a later call—sometimes after application logic validates or transforms it.
A simple prompt chain
For example, imagine analysing customer feedback.
Call 1 — Extract
Extract the key points.
The extracted points become the intermediate result for the next call.
Call 2 — Analyse
Analyse those points and identify the main issue.
That analysis becomes the intermediate result for the final call.
Call 3 — Recommend
Based on the analysis, generate a recommendation.
These are three separate LLM calls. They can all use the same model.
Separate calls, not steps inside one prompt
This is different from putting Step 1 → Step 2 → Step 3 inside a single prompt. Those instructions still run within one LLM call.
Prompt chaining introduces an explicit boundary between calls. That boundary allows an intermediate result to be inspected, tested, transformed, or changed before the workflow continues.
When prompt chaining helps
Prompt chaining is useful when an intermediate result needs independent handling. If the stages do not need to be inspected or changed separately, one well-structured call may be the simpler option.
The trade-off is that every additional call adds latency, cost, orchestration, and another opportunity for an earlier error to propagate.
A building block for larger workflows
A prompt chain does not have to exist on its own. A Plan → Execute → Verify workflow, for example, may contain one or more prompt chains when its stages use separate coordinated LLM calls.
You don’t always need one giant prompt. You also do not need a chain for every task.
The useful design choice is where to introduce boundaries. When an intermediate result deserves its own point of inspection or control, prompt chaining gives that boundary a clear place in the workflow.