Claude Certified Engineers logoCCE

CCAO-F · Study Guide

← Domain 1: Prompting and Task Execution

1.2 · Lesson 2 of 4

Task Decomposition

What you need to know

Compound requests fail quietly in claude.ai because Claude has to juggle several goals in one pass. Splitting the work into a chain of smaller prompts, each with a checkable output, gives you a place to inspect and correct before errors carry forward. This is task decomposition for associate-level product work: research, then filter, then draft — not agent orchestration.

When to decompose

Decompose when a request bundles different jobs, such as:

  • Read source material and extract facts
  • Analyse or filter those facts
  • Draft an email, brief, or checklist from the filtered set

Keep a single prompt when the ask is already one clear deliverable with enough context.

Prompt chaining with checkable steps

Design each step so its output can be read on its own: a bullet list of decisions with quotes, a table of options, an outline. Feed only the approved output into the next prompt. A wrong outline is cheap to fix; a wrong 3,000-word report is not.

  1. 1Extract — pull the facts or obligations from the source.
  2. 2Check — you (or a dedicated review prompt) verify that list.
  3. 3Produce — write the summary, email, or checklist from the checked list.

Self-correction chains — and their limit

A useful pattern is: draft → ask Claude to compare the draft to stated requirements → list gaps → revise. That catches missing sections and inconsistent tone. It does not replace comparing consequential claims to the real document, because the same model is judging its own work.

Worked example

Meeting notes are messy. One prompt asks Claude to extract decisions and draft the follow-up email at once. The email announces two decisions the team never made.

Fix: first ask for a numbered list of decisions with a supporting quote from the notes. Verify that list. Only then ask Claude to draft the email from the verified list, with a rule not to add decisions that are not on it.

Exam traps

  • One giant prompt is fine if you tell Claude not to skip anything

    Emphasis does not create checkpoints. Decomposition does.

  • A self-correction step proves the chain is factually correct

    Asking Claude to review its own draft against criteria helps clarity and consistency. It is not a substitute for checking claims against the source of record.

  • Decompose every request, including trivial one-step asks

    Extra steps add effort without quality when the task is already a single checkable job.

  • Drafting and extracting decisions can share one unreviewed pass

    Invented decisions sneak into emails when extraction is not verified first.

Practice scenario

A compliance analyst sends one message: summarise this long regulation, find which duties affect HR, draft an email to managers, and build a checklist. The reply is shallow and the email includes duties that do not apply to HR. What is the best next move?

Choose one answer

Build exercise

Split a Compound Request Into a Prompt Chain

1.2 · Intermediate · ~30 min · claude.ai

What you will practise

  • Why one prompt that asks for five things tends to do some of them poorly.
  • How to break a request into steps you can check one at a time.
  • How the output of one step becomes the input of the next.
  • How a self-correction step can catch errors before you rely on them.

Full steps, hints, and the skill check live on the domain exercises page. Keep this lesson open while you work so you can apply what you just studied.

Open Domain 1 exercises

Path: /learn/claude-associate/exercises/1-prompting-task-execution

Sources

View progress