Chain-of-Thought Prompting
Tier: Foundational · Task types: Reasoning, maths, logic, multi-step analysis · Difficulty: ⭐⭐☆☆☆
Table of contents
Definition
Chain-of-thought (CoT) prompting instructs the model to reason through a problem step by step before giving the final answer. Instead of jumping to a conclusion, the model externalises its reasoning process, which makes errors visible and correctable — and dramatically improves accuracy on tasks that require multi-step logic.
The simplest CoT trigger is appending “Think step by step” to your prompt. More powerful versions provide an example of step-by-step reasoning (few-shot CoT).
Use Cases
- Mathematical word problems where intermediate calculation steps are needed for accuracy
- Multi-condition logic — decisions that depend on several factors being weighed in sequence
- Argument analysis — identifying premises, assumptions, and conclusions in complex texts
Weak Prompt
A store sells apples for £0.40 each and oranges for £0.65 each.
Sarah buys 7 apples and 4 oranges. She pays with a £10 note.
How much change does she receive?
Weak Prompt Output (Gemini)
Sarah receives £3.60 in change.
Assessment: Wrong. (Correct answer: £10 − (7×£0.40 + 4×£0.65) = £10 − (£2.80 + £2.60) = £10 − £5.40 = £4.60.) The model jumped to an answer without showing its working.
Improved Prompt
A store sells apples for £0.40 each and oranges for £0.65 each.
Sarah buys 7 apples and 4 oranges. She pays with a £10 note.
How much change does she receive?
Think through this step by step before giving the final answer.
Improved Prompt Output (Gemini)
Step 1: Calculate the cost of apples: 7 × £0.40 = £2.80
Step 2: Calculate the cost of oranges: 4 × £0.65 = £2.60
Step 3: Calculate the total cost: £2.80 + £2.60 = £5.40
Step 4: Calculate the change: £10.00 − £5.40 = £4.60
Sarah receives £4.60 in change.
Assessment: Correct, clearly reasoned, auditable. The step-by-step format also makes it easy to spot where an error would occur if the numbers were changed.
Verdict
Chain-of-thought is one of the highest-impact techniques for reasoning tasks — the improvement in accuracy is consistent and significant. The trade-off is output length: CoT responses are longer and may require post-processing to extract the final answer if parsing programmatically.
Common Pitfalls
1. Applying CoT to simple tasks “Think step by step” on a basic factual lookup adds verbosity without benefit. Reserve CoT for genuinely multi-step problems.
2. Trusting the reasoning but not verifying the answer CoT makes errors visible, but the model can reason convincingly toward a wrong conclusion. Always verify final answers on high-stakes tasks.
3. Not extracting the answer cleanly If you’re parsing CoT output programmatically, add an explicit instruction like “End your response with: ANSWER: [value]” to make extraction reliable.