Prompt chaining

What is it?

Prompt chaining gives predefined tasks to separate model calls in sequence. The output of one step becomes the input of the next. A check at each handoff prevents a wrong intermediate result from being carried through the whole chain.

When does it help?

In repeated work where stages such as extraction, drafting and formatting are clearly separated.

Examples

The situations and responses below are fictional teaching examples; they are not results from a model, tool or benchmark that was actually run.

Simple

Situation

You are preparing a workshop announcement in two steps.

Prompt

Initiator: A human; two separate chat calls with a human check in between.
Call 1: From this note, extract only the definite information: "Drawing workshop, 12 October, 14:00, 8 people. Venue to be announced later." Write the fields date, time, capacity, venue.
Check: Confirm that the four fields match the source; if the venue is unknown, keep it as empty information.
Call 2: Using the checked fields, write a two-sentence announcement. Do not add a new venue, fee or registration link. Finish if the final text keeps the four fields.

Sample output

Intermediate output: “12 October; 14:00; 8 people; venue not announced.” Final output: “The drawing workshop will take place on 12 October at 14:00 with 8 places. The venue will be announced later.”

What did we get?

The information stayed fixed while the writing style changed.

Medium

Situation

You are turning three customer reviews into an anonymous summary.

Prompt

The application should run 3 separate calls; carry only the previous approved output.
Input: "Ada: Setup is easy." "Bora: Setup is easy but the text is small." "Cem: The text is small."
1. Remove the names; give each review the ID y1, y2, y3. Human check: if a name remains, stop.
2. From the anonymous reviews, extract themes and the IDs of the supporting reviews. Every ID must exist; if not, try one correction, then stop.
3. From the theme list, write a two-point product summary. Do not add causes beyond frequency, or generalizations to all customers.

Sample output

“Easy setup: y1, y2. Small text: y2, y3.” Then: “Two of the three reviews mention easy setup, and two mention small text.”

What did we get?

Each stage did a different job; personal data did not reach the final writing call.

Hard

Situation

A release draft will be produced from document changes, but if there is a conflict, the chain must stop.

Prompt

Supervisor state object: documents, extracted_changes, conflicts, draft. At most 4 model calls; no publishing tool.
Input A: "v2: Export to CSV and JSON." Input B: "v2: Only CSV is supported."
1. Extract each claim with its A/B ID.
2. Check for disagreement about the same feature. Do not assume either source is authoritative.
Handoff gate: if conflicts is not empty, do not run the writing call. Show the human only which information needs to be chosen.
If the human separately reports that source B is valid, add this decision to the state record; in call 3, write the release draft only from the approved information. Stop at draft generation.

Sample output

“JSON support is contradictory: A says yes, B says no. The draft stage was not entered before the authoritative record was determined.”

What did we get?

Instead of a fast chain, we got a checkable handoff.

Where should you stop?

Opening a separate call for every paragraph may be unnecessary. The value of the chain lies in dividing the work and checking the handoffs. ReAct chooses its path based on observations; in this method the path is known in advance. Keeping calls separate does not, on its own, provide independent verification.

Sources

• Building Effective AI Agents \ Anthropic (new tab) — Anthropic. 2024-12-19. Explains the prompt chaining pattern of splitting work into predefined subtasks checked through intermediate gates; a provider engineering guide. Evidence level: page body.

  Copy source address