How to Get AI to Think Through Complex Problems

Let me ask you something…

Have you ever given your AI a really complicated problem and just hoped it would figure it out?

And what happened?

It probably spit out some half-baked answer that didn’t really help.

Either the answer was too short and didn’t cover what you wanted, or it was off, meaning it was irrelevant.

Or both.

The point is: If you want AI to tackle complex problems, you need to guide it step by step.

It’s because AI “feeds” on the context you give it. It steers its internal probabilities for the next words into the right direction.

I’ve made a video about this—watch it if you want to learn more about how LLMs work, to improve your prompting.

But back to my point: that’s where Chain of Thought (CoT) prompting comes in.

Let me break it down: Instead of asking either one big, broad question, you ask a series of smaller, open-ended questions to build up the context.

It’s like walking the AI through your thought process. This way you let it fill up its own context.

For example: Let’s say you’re comparing email providers and trying to figure out which one has the best deliverability for your cold email campaigns.

You don’t just ask, “Which provider is best?”, because you’ll get a very small part of the answer.

Instead, you build up the question: First, ask about what criteria to pay attention to in general when choosing a provider.

Then, ask about what providers are on the market.

Finally, get specific—like “How does [provider xyz]’s compare to others?”

It’s like a funnel, in that you narrow down the possible answers of the AI towards what you actually want to know.

See what’s happening? 

You’re feeding the AI more and more context, and by the time you ask the final question, it’s got everything it needs to give a thoughtful, accurate response.

Ultimately it’s about constraining the possible output choices of the AI down to exactly the topic/question you want.

You ask the actual question at the very end!

Prompting is all about constraining. The more context you give, the more constrained the AI is in order to adhere to your request.

It’s like buying shoes.

You don’t just ask the shop assistant to give you “shoes”. If you do, they’ll come back with the first pair they see, because ANY pair satisfies your request—all of the hundreds of pairs they sell!

But if you say you want red shoes, then there’s maybe only fifty candidates.

If you say you want something for the summer, then it’s maybe twenty.

When you mention your foot size, you reduce the set to five.

Now you can go through all five and choose the one you like best.

Prompting is no different. If you can ask for something in real life, you can also write prompts!

Let’s see ChatGPT’s answers in the mailbox comparison example. I first asked it to output the comparison criteria for mailboxes.

This detailed table mentions most of the things I need to consider. Some I didn’t even know about!

This is important, because it educates me on the problem, and it gives enough context breadth to answer my next questions. It sets the stage.

Now I want to see the providers on the market.

Because we’ve already “talked” about criteria with the AI, ChatGPT has no problem “remembering” them and showing me this nice table.

Notice the shorter prompt I gave it? Win-win: I can save time writing and ideate quicker.

Lastly I cat get to the actual question: I want to know how a specific email provider compares in terms of deliverability.

Basically I zoom into one single field of that table.

The answer is longer than that, but we’ll keep it here for simplicity. At this point ChatGPT gave me a comprehensive answer regarding my initial question. 

I won’t be digging any deeper, you’ve got the point by now. If you’re curious, you can see the complete conversation right inside ChatGPT.

It’s a game-changer.

Once you learn how to use CoT prompting, you can tackle any complex problem with ease.

Want to see how CoT works in other scenarios?

I’ve got an entire chapter dedicated to Chain-of-Thought in the course inside the Make Work Obsolete community.

Click the link below and start solving complex problems today.

👉 Join Make Work Obsolete and master prompting today

Cheers,
Robert