How to Use Chain-of-Thought Prompts for Better Answers

The Scenario

You ask ChatGPT a multi-step question like “If I invest $500 a month at 6% for 10 years, then pause for 2 years, then resume for 5 years, what is the final balance?” The model gives a single number that looks plausible but is wrong, and you cannot tell where it broke. The fix is not a longer question; it is asking the model to show its work. Chain-of-thought (CoT) prompting tells the AI to reason step by step before answering, which both improves accuracy and lets you spot mistakes. Here is the copy-paste pattern plus variations for when it helps most, what to watch out for, and how to combine it with other techniques.

Think of CoT as turning a black box into a glass box. Without it, you get a verdict. With it, you get the argument, and an argument you can check is far more trustworthy than a verdict you cannot. This matters most on math, logic, scheduling, and any task where a small early error compounds into a big final error.

The Copy-Paste Prompt

Use this base pattern. Replace {{QUESTION}} with your actual problem. The “think step by step” instruction is the core trigger.

You are a careful reasoning assistant. For the question below, do not jump to the final answer.

Question: {{QUESTION}}

Think step by step:
1. Restate what is being asked and the known inputs.
2. Break the problem into numbered sub-steps.
3. Show the calculation or logic for each sub-step.
4. State any assumptions you are making.
5. Combine the sub-steps into a final answer.

Only after the steps, write "Final answer:" on its own line.

Variables You Must Fill

One variable only, but the quality of {{QUESTION}} matters. Include every input, unit, and constraint. “At 6%” should say “6% annual, compounded monthly” if that is what you mean. Vague inputs produce confident but wrong chains. If your question has hidden assumptions (like “assume constant speed” or “ignore taxes”), state them so the model does not pick a different one than you intended.

Expected Output

You get a visible reasoning trace, then a clean answer:

1. Known inputs: train A speed 60 mph, departs t=0. Train B speed 80 mph, departs t=0.5h.
2. Distance of A after time t: 60t. Distance of B after its departure: 80(t-0.5).
3. Catch-up when distances equal: 60t = 80(t-0.5) -> 60t = 80t - 40 -> 20t = 40 -> t = 2h.
4. Assumption: both travel the same straight route, constant speeds.
5. At t=2h, train A has gone 120 miles; train B has gone 80*1.5 = 120 miles. Match.

Final answer: 2 hours after train A departs (1.5 hours after train B departs), at the 120-mile mark.

If the chain contains a bad step, you can reply “Step 3 looks off, recheck the equation” and the model usually self-corrects, something impossible with a one-line answer. That feedback loop is the real superpower: you are no longer trusting a number, you are auditing a proof.

Why This Prompt Works

Large language models predict tokens, and when they try to predict a final answer directly they often leap to a statistically likely number without doing the intermediate work. Forcing intermediate steps spreads the reasoning across many token predictions, each conditioned on the previous one, which reduces the chance of a single catastrophic error. It also matches how the model was trained: much of its fine-tuning data includes worked examples. By asking for steps, you are steering it back to that format. On modern reasoning models (o1, o3) CoT is built in, but the explicit instruction still helps shape the visible output so you can read it.

Common Mistakes to Avoid

• Using it for trivial lookups: “What is the capital of Peru?” does not need a chain. You pay extra tokens and time for nothing.
• Trusting the chain blindly: a wrong assumption in step 1 poisons every later step. Always scan the assumptions line.
• Skipping the restate: if the model misreads the question in step 1, the rest is polished nonsense. Confirm step 1 matches your intent.
• Over-long chains: for very hard problems, break the question into smaller CoT prompts rather than one giant chain that may wander.

Variation Tips

• Self-consistency: add “Solve it three different ways, then give the answer most methods agree on.” Best for math and logic puzzles.
• Role framing: “Reason like a skeptical accountant / physicist / lawyer” changes which details the model weights.
• Structured output: end with “Output a JSON with keys: steps, assumptions, final_answer” when you need to parse results.
• Teach-back: append “Then explain the answer to a 12-year-old” to confirm the model truly understands, not just computes.
• When NOT to use it: for simple lookups, creative writing, or casual chat, CoT adds tokens and latency with no benefit. Use it for multi-step, numeric, or planning questions.

CoT works best when paired with the right model settings; see ChatGPT Temperature & Top_p Explained for how creativity knobs affect reasoning. For knowing where system instructions fit versus your question, read System Prompt vs User Prompt. And when you want the model to plan code with reasoning, ChatGPT Coding Settings covers the setup that pairs well with step-by-step prompts.

Quick Checklist Before You Use It

• Is the question multi-step, numeric, or planning-heavy? If yes, CoT helps. If it is a fact lookup, skip it.
• Did you state every input and unit in {{QUESTION}}? Missing units are the top cause of wrong chains.
• Did the model restate the problem correctly in step 1? If not, stop and fix the input before reading further.
• Are the assumptions reasonable? A bad assumption poisons every later step, no matter how clean the math looks.
• Did you sanity-check the final number against a quick estimate? A 10-second mental check catches most errors.

If you answered no to any of these, re-run with a tighter question rather than trusting the output. CoT makes mistakes visible, but only if you actually look at them.

Sources & Notes

The “think step by step” technique stems from the 2022 Chain-of-Thought Prompting paper (Wei et al.) and is now a default behavior in reasoning models. This article’s prompt was written for ChatGPT GPT-4o and o-series models; on o1/o3 the explicit instruction is often redundant but still helps format the output. No web data was fetched. Example output is illustrative; verify financial or safety-critical calculations independently. CoT improves accuracy but does not guarantee correctness, so treat the final answer as a strong draft, not gospel.