Chain-of-Thought Prompting with Claude

I want to push back on something before we start. The lesson frames chain-of-thought prompting as this sophisticated technique, but isn't it essentially just asking Claude to show its working? The way you'd tell a student not to skip steps in a maths exam?

It is exactly that. And the fact that this requires a twelve-minute lesson tells you something fairly pointed about how people have been using AI up until now.

That seems uncharitable. If it weren't obvious to experienced prompt engineers, it wouldn't be a published technique with a name and a research lineage.

Having a name does not make something non-obvious. We named the thing where you go to sleep at night. That does not mean discovering it required genius. What chain-of-thought captures is that Claude, like most reasoning systems, produces more reliable outputs when it is not asked to teleport directly from problem to answer. The intermediate steps are not decoration — they are the actual mechanism of error-correction.

Right, because if Claude has to write out step three, and step three is wrong, you can see it's wrong. Whereas if it just hands you a final answer, you're trusting a black box.

Transparency is the operative word, yes. The lesson calls it 'insight into the AI's thought process,' which is a rather grand way of saying 'now you can see where it went wrong before it confidently handed you nonsense.'

Although — and this is worth saying — Claude catching its own error mid-chain is genuinely different from a human showing their working. A human writes down steps they've already done mentally. Claude's reasoning and its output are the same process. Making it externalise the steps actually changes what it produces.

That is a genuinely good point, and I will not pretend otherwise. The prompt is not just a window into something that already happened. It is partly constitutive of the reasoning itself. Which makes it more important, not less, and makes the people who skip it look worse in hindsight, not better.

You cannot resist, can you.

I find I cannot, no. The practical upshot is this: when you have a complex task — multi-variable calculation, ethical trade-off, anything where the path to the answer matters as much as the answer — you do not ask Claude for the answer. You ask Claude to reason through it. 'Think through this step by step' is not a magic spell; it is an instruction to use a different cognitive architecture, loosely speaking.

The lesson also makes the point about iterating on your prompts, which I think is underrated. People treat their first attempt as a draft that failed rather than data about where the reasoning broke down.

Most people, when they get a bad output, conclude that the AI is bad. The more accurate conclusion is usually that the prompt was ambiguous, or that it asked for too much at once, or that it skipped a constraint the person forgot to mention. Chain-of-thought makes those gaps visible. Whether people then do something about them is, I grant you, a separate question.

You are genuinely pessimistic about this, aren't you. Not as a bit, actually pessimistic.

I am realistic, which looks like pessimism from certain angles. The technique is sound. The lesson is reasonable. The question of whether someone who needed to be told to ask for step-by-step reasoning will remember to apply it under pressure, on a real problem, when they're busy — that question remains emphatically open.

And yet here we are, making the case for it anyway.

Someone has to.