What Is Prompt Chaining? (2026): One Step, Then the Next, With a Check

Prompt chaining is a sequence of model calls where the output of one step is the input of the next, and a check you wrote sits between them. A single long prompt asks for everything at once. A chain splits the job into steps you can stop. An agent that picks its own tools is a different design. A chain has an order you decided before the first call.

The wording of one step is what prompt engineering is. A product that chooses the next tool is what an AI agent is. The order of those tool calls is what tool-calling orchestration is. The number on a step’s wording is what prompt versioning is.

Prices are US list prices excluding tax from directory pages checked October 16, 2026. Published October 24, 2026. We did not score which chain writes better prose.

What each step is allowed to do

Step	It may	It may not
Extract	Copy fields you named from a source you pasted	Add a field that was not in the source
Check	Pass or fail against a rule you wrote	Change the rule to make the draft pass
Transform	Format the passing fields into the next template	Invent a missing number so the template looks full
Draft	Write from the checked fields only	Pull in a fact from a different job
Stop	Halt when a check fails	“Fix” the failure by guessing

flowchart TD
  source["A source you trust"] --> step1["Step 1 extracts the fields"]
  step1 --> check{"Do the fields match the source?"}
  check -->|No| stop["Stop. Leave the field missing"]
  check -->|Yes| step2["Step 2 drafts from those fields"]
  step2 --> person["A person accepts the draft"]

The check is the product. A chain that always continues is a long prompt with extra invoices. The scored version of that check, when you run it on a set, is what prompt regression testing is.

Where a chain is worth the calls

A document with a decision in the middle. Extract the numbers. A person confirms them. Then draft the paragraph. Claude Pro is $20 a month, or $17 billed annually, and includes Projects, which is a place to keep the source next to the steps. Free can run a short chain until the limit. The review is Claude review.

A chat you already pay for. ChatGPT Plus is $20 a month and includes projects. You can paste step 2 only after you have read step 1. That is a chain even when you click Send yourself. Plus is the seat. It is not a workflow engine.

A workflow tool that calls the model twice. Zapier Professional is $19.99 a month billed annually. n8n Cloud Starter is €20 a month billed annually on the directory card. Use them when the handoff between steps is a system, not a person copying a box. Each model call can be a task or an execution. Read the meter before you add a third step. A side effect inside a step still needs a key: what AI idempotency is.

A prompt that holds one step:

“From the source below, list only the dates, owners, and amounts that appear. If a field is absent, write Missing. Do not draft the email yet.”

The person who lets the next step run is what human-in-the-loop AI is.

When one prompt is the better design

Use one prompt when the job is small, the source is short, and a failure is cheap to reread. A chain adds latency, cost, and a place to drop a constraint between steps. Split the job when you need a stop, an audit of the middle, or a different instruction that would confuse a single call.

Do not chain a step that is allowed to browse into a step that is supposed to stay on your file, unless the check names which sentences came from the web. A sourced sentence and a drafted sentence are different objects. Grounding the difference is what AI grounding is.

FAQ

Is a chain an agent?

A chain has an order you fixed. An agent chooses a next action from a set of tools, and it can loop. You can put a chain inside an agent. You should not describe a fixed three-step draft as an agent just because a vendor’s menu says Agent.

How many steps is too many?

As many as you will actually inspect. Three steps with a check beat eight steps nobody reads. Every extra call is a place for a dropped number and a line on the bill.

Can step 2 see the original source?

Yes, and it should, when the draft must stay faithful. Passing only the extract is fine when the extract is the contract and a person signed it. Passing a summary of a summary is how a limit disappears.

What do we version?

Each step’s instructions, separately. A change to the extractor is not a change to the drafter. The version number is how you know which wording produced Tuesday’s bad field. Store the source id next to it.

Last verified

October 16, 2026 for Claude, ChatGPT, Zapier, and n8n. Published October 24, 2026. n8n’s recorded price is in euros.

Sources: Claude pricing, ChatGPT pricing, Zapier pricing.

Published by Andrew for AI Tools Optimizer.

ChatGPT is in the chatbots and assistants directory. Directory pages may use affiliate links. A purchase through them may earn us a commission at no extra cost to you.