Chain-of-thought prompting: Teaching an LLM to ‘think’
You can make chatGPT better by teaching it how to think like you!
You ask ChatGPT:
“What’s 27 × 43?”
It answers in a second:
“1,161.”
Great. But now ask:
“Let’s solve 27 × 43 step-by-step.”
This time, it replies:
“First, 20 × 40 = 800. Then 20 × 3 = 60. Next, 7 × 40 = 280. And 7 × 3 = 21. Add them all up: 800 + 60 + 280 + 21 = 1,161.”
Same answer. But now… it makes sense.
That’s the magic of Chain-of-Thought prompting.
It doesn’t just give you the final answer… it helps the model (and you) think through the steps to get there.
And it turns out, in more complex, logical, or mathematical instances, that makes a huge difference.
🤖 LLMs Don’t “Know” Things. They Predict Them
Let’s remind ourselves how LLMs like ChatGPT work.
In my previous post, I’ve explained how they’re trained on massive amounts of text and optimized to predict the next word in a sentence… not to recall facts, and definitely not to reason like a human.
So when you ask a question, they don’t go fetch an answer from a database. They try to generate what a reasonable-sounding answer should look like.
That’s why LLMs can ace trivia questions, but when it comes to multi-step logic or math, it can start hallucinating and giving you all sorts of weird answers!
… unless we prompt them differently.
💡 What Is Chain-of-Thought Prompting?
Chain-of-Thought (CoT) prompting is a simple approach to prompting, but with a big impact:
Instead of just asking for the answer, you ask the model to show its work.
For example:
❌ “What’s the fastest way to reduce churn?”
✅ “Think step-by-step about why users churn, then suggest three retention strategies based on that reasoning.”
This small shift encourages the model to break down a problem, not just guess the end result.
It’s the difference between:
A student blurting out an answer on a test
vs.
A student explaining how they got there
In LLMs, that explanation often leads to a more accurate, more useful output.
📈 The Research Behind CoT (And Why It Works)
This isn’t just a hunch, it’s backed by serious research.
A 2022 paper titled Chain-of-Thought Prompting Elicits Reasoning in Language Models tested this ‘chain-of-thought’ approach that we discussed across tasks like:
Grade-school math (GSM8K)
Logic puzzles
Commonsense reasoning
Symbolic math
And the results were dramatic.
For example, on GSM8K:
Standard prompting: 18% accuracy
Chain-of-Thought prompting: 57% accuracy
(this was using PaLM 540B, one of Google’s largest models at the time)
That’s not a small bump. That’s a 3x improvement.
And the bigger the model, the more it benefits. CoT prompts help unlock latent reasoning ability in large models that otherwise sit unused.
🧠 Why Chain-of-Thought Works So Well
The trick behind CoT isn't just formatting… it's cognitive scaffolding.
Here’s what’s happening:
LLMs are trained on tons of step-by-step explanations (think: Stack Overflow, Wikipedia edits, math tutorials).
CoT prompting activates those patterns by mirroring them.
By asking the model to “think step-by-step,” you’re nudging it toward more grounded, logical paths.
It’s not that the model becomes smarter. It just becomes more careful, more deliberate.
You’ve given it permission to reason.
🛠️ How to Use CoT in Your Everyday Prompts
Chain-of-Thought isn’t just for math problems.
It works anywhere you need structured thinking.
Here are some examples:
📊 1. Strategy or Problem Solving
Instead of:
“What should our go-to-market plan be?”
Try:
“Let’s walk through the go-to-market plan step-by-step. Start by identifying our audience, then messaging, then acquisition channels.”
Why it works: You’re nudging the model to break down a complex decision (not just spit out a generic strategy).
✍️ 2. Writing and Ideation
Instead of:
“Write a blog intro.”
Try:
“List out three different ways to open a blog post on this topic… one with a story, one with a stat, and one with a provocative question. Then explain which one would be most effective.”
Why it works: You’re activating lateral thinking, not just surface-level phrasing.
🎓 3. Learning and Explanation
Instead of:
“Explain quantum physics.”
Try:
“Explain quantum physics to a 12-year-old, step-by-step, starting from atoms and working toward uncertainty.”
Why it works: You’re forcing the model to build knowledge from the ground up.
🧠 Does Training Models With CoT Make Them Reason Better?
Yes! And it’s not just about the output. CoT-style reasoning actually teaches the model to internalize better logic paths.
Later studies found that:
CoT-trained models generalize better to unseen problems
They’re more robust to ambiguity
And they make fewer glaring reasoning errors
Even more interesting:
Once a model is “exposed” to CoT patterns, it tends to use them even when you don’t explicitly ask.
It’s like teaching a student to show their work… eventually, they just start doing it.
⚡ Try This: A Quick CoT Prompt You Can Use Today
“Think step-by-step about why some projects succeed while others don’t. Then turn that into 3 key lessons for project mangers.”
Watch what the model does.
Then try:
“Now reverse those lessons: what causes projects to fail?”
This layered reasoning is where CoT really shines.
🔁 Final Thought: Don’t Just Ask for Answers. Ask for Thinking.
In a world where speed is everything, Chain-of-Thought prompting invites us to slow down… just enough to think more clearly.
For the AI, it leads to better outputs.
For us, it builds trust and clarity.
And when done right, it turns a black-box machine into a transparent thinking partner.
So the next time you’re stuck?
Don’t just ask ChatGPT for a solution.
Ask it to walk with you. Step-by-step.
Next on LLMentary:
Now that we’ve learnt how to make an LLM ‘think’, we’ll next understand how to make them ‘reason’…. using a technique AI researchers call ‘Enhanced’ Chain-of-Thought. I’ll be covering that in my next post.
If you want to be notified when I drop that, I’d really appreciate it if you could subscribe to LLMentary!
Until then, stay curious.
Share this with anyone you think will benefit from what you’re reading here. The mission of LLMentary is to help individuals reach their full potential. So help us achieve that mission! :)
Share LLMentary to win rewards! When you use the referral link below, or the “Share” button on any post, you'll get credit for any new subscribers. Simply send the link in a text, email, or share it on social media with friends.
I’m trying different writing and understanding styles, such as quick and easy, dense and informative, and technical and applicative. Feel free to leave a comment on how you prefer your information, and I’ll incorporate your feedback into my writing style as well!