How does it differ from least-to-most prompting?

Least-to-most is a particular decomposition-and-solving strategy that can be implemented as a chain. Prompt chaining also covers extraction, drafting, checking and other sequences.

Can stages run in parallel?

Independent branches can. A stage that needs a previous result must wait for that dependency.

Does the model choose the next stage?

In a fixed chain, application logic usually does. More dynamic systems may let a model choose actions, but that is an additional design choice.

AI Glossary / Prompting and human-AI interaction / Intermediate

Prompt chaining

Definition

Prompt chaining connects multiple model calls so that an earlier call's output becomes part of a later call's input. Each stage handles a defined part of the task, and the application passes selected results between stages. A chain can also include validation, tools or human review.

What it means

Instead of asking one prompt to extract facts, analyze them and draft a report simultaneously, an application can organize those activities as separate steps. The extracted facts feed the analysis, which then informs the report.

The connection between steps is what makes a chain. Unrelated sequential prompts do not form this dependency. Independent prompts can also run in parallel, but that is a different relationship.

Chaining offers places to inspect intermediate results and diagnose failures. It also creates dependencies: an omitted fact in the extraction stage can disappear from every later answer, even when the final prose is fluent.

How it works

Design each stage around an explicit input and output. For example, an extraction stage returns structured fields, a validation stage checks required values, and a drafting stage converts validated facts into a reader-friendly summary.

The application stores the relevant state and constructs the next prompt from selected fields. It need not copy the entire conversation every time. Retaining links to original evidence allows later checks.

A failed check can stop the chain, request correction or route the case to a person. Retry rules should specify what is repeated and what is preserved. The same model can serve every stage, or different models and tools can be used. Chaining describes the workflow relationship, not a requirement for multiple autonomous agents.

Three practical examples

A publishing workflow extracts claims from a draft, checks them against supplied evidence and then rewrites the supported text for a particular audience.
A customer-service workflow classifies a request, retrieves the relevant procedure and drafts a reply grounded in that procedure.
A research workflow extracts structured observations from separate documents and then produces a comparison from the checked records.

Think of it this way

Think of an editorial process in which notes become an outline, the outline becomes a draft and the draft receives a review. Each handover has a purpose and an expected form.

The analogy highlights why handovers matter. If the notes omit a qualification, later polishing will not automatically restore it. A useful chain preserves enough evidence to check whether the final result still reflects the source.

Worked example

Consider an original two-stage service-request example. Stage one extracts issue type and customer reference from a message. Stage two uses those fields to select an approved response template and draft a reply. The final answer is considered successful only if the extraction is correct and the reply correctly follows the selected procedure.

In a teaching population of 100 requests, suppose stage one extracts both required fields correctly for 90 requests. Among those 90, stage two produces a correct reply for 81. Assume none of the ten incorrectly extracted requests produces a successful final answer.

End-to-end success is 81/100 = 81%. Stage two's conditional success rate is 81/90 = 90%. The product 0.90 × 0.90 = 0.81 is valid here because the second figure is explicitly conditional on correct extraction; it does not require an independence assumption.

It would be wrong to take two unrelated 90% benchmark scores and automatically multiply them. Their evaluation conditions may differ, and downstream behavior can depend on the type of upstream error.

Now add an extraction check that identifies five of the ten bad records and routes them for review. Those five have been detected, not automatically corrected. A report should distinguish automated success, review workload and final outcomes after review. The counts are invented.

Illustrative case study

Fictional case: Pine Support builds a chain for warranty enquiries. It first extracts the product code, then retrieves the matching policy, then drafts a response. A check prevents drafting when the code is missing or no policy is found.

Reviewers discover that a valid-looking code can still be the wrong code. They add source-location information to extracted fields so a person can inspect the original message when needed. They also preserve the policy version used to prepare each draft.

The team compares the complete chain with a simpler baseline and examines failures at each handover. This fictional case claims no measured accuracy improvement.

Common misconceptions and mistakes

More stages guarantee higher quality: every additional handover can introduce omissions or errors.
A valid structure means valid facts: a perfectly formatted product code can still be incorrect.
A critic model is an independent source of truth: it may share the first model's weaknesses or lack the evidence needed to check it.
Prompt chaining requires multiple agents: an ordinary application can run a fixed sequence of calls.
Retrying the whole chain is always harmless: workflows with external actions must avoid repeating effects such as duplicate record creation.

Business use

Chaining is useful when a task has separable stages with meaningful intermediate checks. Teams can revise one stage and inspect its effects.

The trade-offs include additional latency, input and output processing, and more complex failure handling. Measure complete-task success rather than celebrating isolated stage scores. Keep the original evidence accessible, version the prompts and define what happens when a stage cannot produce a trustworthy result. For simple transformations, a single prompt or conventional code may be easier to operate.

Check your understanding

A downstream stage is 95% correct when given perfect inputs. Does that establish 95% success for the full chain?

Answer: No. The full workflow also includes the quality of upstream outputs and how downstream processing behaves when those outputs are incomplete or wrong.

Related terms

Task decomposition
Least-to-most prompting
Validation
Orchestration
Error propagation
Workflow
structured output
intermediate state
retry
human review

Learn next

Draw a proposed chain and label each handover with its required fields and evidence. Choose one failure case per stage. Test whether the workflow stops, corrects or escalates appropriately, and compare its end-to-end result and cost with a simpler baseline.

Sources

LangChain, Workflows and agents, prompt-chaining pattern and workflow distinctions: https://docs.langchain.com/oss/python/langgraph/workflows-agents

Zhou and colleagues, Least-to-Most Prompting Enables Complex Reasoning in Large Language Models, 2023, for a related sequential prompting method: https://arxiv.org/html/2205.10625v3

The conditional-rate calculation and Pine Support case are original hypothetical illustrations.

Previous: Prompt-based learning Next: Prompt engineering