Map
Framework

Prompt chain with per-step checks

The question it answers

When a marketing team asks AI to produce longer content such as a post or a video script, into which steps should the job be split, and where should a person sign off to keep quality under control?

Origin
Anthropic, 2024
Diagram type
Process by phase
Level of analysis
Individual
Topics
AI for leadersMarketing & sales
Glass sphere with a glowing neural-network core, icon for AI for leaders

Diagram

Prompt chain with per-step checks

  1. Four example steps
  2. Analyse the audience and contextAfter each step: accept, refine or discard.
  3. Generate 3 to 5 angles; the human choosesAfter each step: accept, refine or discard.
  4. Draft the contentAfter each step: accept, refine or discard.
  5. Adjust the toneAfter each step: accept, refine or discard.

When to use it, and when not to

Use it when

Use it when a task splits cleanly into fixed steps, each taking the output of the one before. (Anthropic, 2024) calls this prompt chaining and says it trades waiting time for accuracy, because each call becomes an easier task.

Avoid it when

Not suited to tasks whose steps cannot be predicted: (Anthropic, 2024) points to orchestrator-workers or agents for those. The four steps are one example for creative content, not a standard set; putting a person at each gate is the site’s practice, while (Anthropic, 2024) speaks only of checks on intermediate steps and pauses for human feedback.

How to apply it, step by step

After each step a person looks at the result and accepts, refines or discards it before the next step. (Anthropic, 2024) allows checks (“gates”) on intermediate steps to confirm the process is still on track; here the gate is a person. Record the version accepted and the reason for each edit so the run can be repeated.

Pitfalls

Pitfall

A chain is longer than a single call and so slower; (Anthropic, 2024) states that this is a trade in waiting time. For a small task one call with examples is often enough.

Sources

References

Listed A–Z by author surname, Harvard style. A source with a DOI links to https://doi.org/…; one without links to its page or says plainly that it has no DOI.

  1. Anthropic (2024) Building effective AI agents. Anthropic Research.No DOIanthropic.comCited in 6 articles