What is Prompt Chaining?
Also called prompt pipelines.
Prompt chaining is the practice of decomposing a task into a sequence of separate model calls, where the output of one call becomes part of the input to the next. Each step handles a narrower subtask than the whole. Chaining trades additional calls and latency for higher reliability, easier debugging, and per-step validation.
Cramming a complex task into one prompt tends to degrade every part of it, since the model must plan, gather, reason, and format simultaneously. Splitting into steps such as extract, then analyze, then draft, then check gives each call one clear objective. Focused prompts are shorter, easier to evaluate, and less prone to instructions being overlooked.
Validation between steps is the biggest practical gain. A structured output from step one can be schema-checked before step two consumes it, so a malformed intermediate result fails immediately at a known place instead of producing a plausible but wrong final answer. Individual steps can also be retried, cached, or routed to differently sized models.
The costs are real. Each additional call adds latency and token usage, and errors compound: a subtle mistake in an early step is often confidently elaborated by later ones rather than corrected. Chains also lose context that a single prompt would have retained, so relevant information must be deliberately threaded forward.
Chaining is a fixed pipeline, distinct from an agent that decides its own next action. Chains have a predetermined shape and are therefore predictable, testable, and bounded in cost. Common patterns include sequential refinement, parallel branches merged by a final call, and a routing step that dispatches to one of several specialized downstream prompts.
Key points
- Splits a task into several focused model calls in sequence
- Allows validation and retry between steps
- Adds latency and token cost per additional call
- Early-step errors get elaborated rather than corrected
- A fixed pipeline, unlike an agent choosing its own steps
In practice
A contract review pipeline runs three calls. The first extracts every obligation into a structured list, which is schema-validated before proceeding. The second classifies each obligation by risk against a rubric. The third drafts a summary memo from the classified list. When the memo is wrong, the team inspects the intermediate outputs and usually finds the fault in extraction, which a single combined prompt would have hidden entirely.