Day 7 — Getting it to think step by step

For anything with more than one step, asking for the working improves the answer.

Why it helps

A model produces text one piece at a time, each piece conditioned on what came before. If it commits to an answer immediately, it has to get there in one leap. If it lays out the steps first, each step becomes context for the next - and the answer at the end is better grounded.

This is not the model "thinking harder" It is not deliberating. Writing the steps produces text that constrains what follows, so the final answer is more likely to be consistent with them. The practical effect is real; the mechanism is simpler than it looks.

The simplest version

Numbering the steps yourself

"Think step by step" on its own helps a little. Naming the steps helps far more, because you are supplying the analytical structure rather than hoping it picks a good one.

Ask for the working, then check it

The steps are what you audit The value is not that the answer is more likely right - it is that a wrong answer becomes visible. If step 2 says "filter to the last financial year" and the query filters on calendar year, you have caught it. A bare answer gives you nothing to check.

Splitting a big task

For anything substantial, several small prompts beat one large one. You inspect the output at each stage instead of discovering a wrong assumption at the end.

Prompt 1: Given these columns, what checks should I run before analysing? Prompt 2: Write the SQL for check 3. Prompt 3: Here are the results. What do they suggest? Prompt 4: Draft two sentences of commentary using only those numbers.

Why this beats one big prompt You keep the judgement. At each step you decide whether to accept, correct or redirect - which is exactly the part of analysis that should stay with a person.

Asking it to critique itself

Models are noticeably better at finding holes in an argument than at avoiding them. Drafting and then critiquing in a separate step is a genuinely useful two-pass workflow.

Try these yourself

Turn "why did sales drop?" into a five-step prompt.
Write a debugging prompt that asks for the cause before the fix.
Split a month-end reporting task into four separate prompts.
Explain why asking for the working makes errors easier to catch.
Ask a model to critique a paragraph of your own analysis.

← PreviousDay 6 — Examples and few-shot prompting Next →Day 8 — AI for Excel