PPROMPT WORKBOOK Build a prompt
Deep dive

Prompt chaining: split complex work into stages

A prompt chain gives each stage one job. Between stages you can inspect, correct or stop before an error compounds.

When is a chain better than one prompt?

Use a fixed chain when the task has known stages and the intermediate result is worth inspecting. Examples include retrieve sources → extract facts → draft → verify claims, or outline → approve → write → edit.

Do not add a chain merely because a task sounds advanced. Every model call adds latency, cost and a new failure point. Start with one call and split when a measured failure can be isolated.

Four rules for a robust chain

  1. One visible goal per stage. The intermediate output must be reviewable.
  2. A narrow handoff format. Pass findings, sources, decisions and open questions, not the entire conversation.
  3. A gate before risky stages. Check format, evidence or approval before continuing.
  4. A clear stop rule. Define when the chain ends or asks a human for help.
Stage 1 — Extract
Find only claims directly supported by the sources.
Deliver: claim | source | source location | uncertainty.

Gate
Stop if an important claim has no source.

Stage 2 — Synthesize
Use only approved rows from stage 1.
Deliver a short decision memo with findings, consequences and open questions.

Chain, workflow or agent?

A chain follows a fixed route. A workflow may contain rules and branches. An agent chooses its next step dynamically. Choose the simplest design that passes the test; see workflows and agents.

Sources