Breaking complex tasks into steps and comparing alternatives

Why break tasks down

A single prompt such as "Plan, write and format a full training course on fire safety" asks the model to make dozens of decisions at once. Quality drops, important steps get skipped and you cannot see where it went wrong. Splitting the work into smaller prompts — sometimes called prompt chaining — gives you:

• checkpoints where you can correct course before errors compound;
• simpler prompts that are easier to test;
• reusable steps (the "outline" step works for many documents).

A typical chain

1. Clarify — "Ask me up to five questions about the audience and constraints."
2. Plan — "Propose an outline with 5–7 sections and one sentence per section."
3. (You review and edit the outline.)
4. Draft — "Write section 2 using the approved outline and these notes…"
5. Review — "Check the draft against these criteria and list problems; do not rewrite yet."
6. Revise — "Fix problems 1, 3 and 4 only."

The human checkpoint at step 3 is essential: it is where your judgement adds the most value.

Passing information between steps

Each step should receive only what it needs: the approved outline, the relevant notes and the criteria. Paste the previous step's approved output into the next prompt rather than relying on the model to remember a long conversation — this keeps each step reproducible and avoids context-window problems.

Comparing alternatives

Single answers hide choices. Ask for options and trade-offs:

• "Give three different subject lines for this email: one formal, one friendly, one urgent. For each, say who it would suit."
• "Compare these two approaches to scheduling volunteers against: fairness, admin time, flexibility. Use a table and finish with which you would choose for a group of 20 volunteers, and why."

Comparison prompts work best when you set the criteria. Otherwise the model chooses criteria that make its preferred answer look good.

Using a model to critique its own output

A useful pattern is generate → critique → revise: ask for a draft, then in a new prompt ask the model to check it against explicit criteria. Self-critique catches some problems (missing sections, broken format) but is unreliable for factual errors, because the same model may repeat the same false belief. Factual checking needs sources (Module 7).

When not to chain

Chaining costs time. For small, low-stakes tasks one good prompt is enough. Chain when the task is large, the stakes are high, or you need to inspect intermediate results.