Prompt Engineering· Lesson 4 of 7
- 1Anatomy of a Prompt — What It's Made Of6 min
- 2Few-Shot — Teach by Example7 min
- 3Structured Output — Make It Return JSON7 min
- 4Chain of Thought — Let It Reason Step by Step8 min
- 5Prompt Iteration — Improving It in Passes6 min
- 6Role & Persona — Tell It Who to Be7 min
- 7Prompt Injection Basics — When Input Turns Against You8 min
Chain of Thought — Let It Reason Step by Step
In one line · For tasks that need reasoning, you ask the model to spell out its steps before the final answer instead of jumping straight to the result.
On multi-step problems the model errs more when it jumps straight to the result; when it 'thinks out loud' its accuracy on math and logic rises noticeably. But it's not magic: on simple tasks it barely helps and can just pad the answer, and modern reasoning models already do this internally without being asked.
Like a student in a math exam: writing the steps reaches the right answer more often than jotting only the final number. Same with the model — though on an easy problem the steps won't change much.
An instant answer is a hunch; chain of thought is working the problem out on paper. Paper cuts mistakes on the multi-step stuff, not on what's already obvious.
Instead of 'how much is left from 200 if I bought a pen for 30, a notebook for 45, and a box of crayons for 65?', say 'first add up the prices, then subtract from 200' — it slips up less.
The model generates its reply token by token, each new token conditioned on everything written before it. When you make it write its steps, the intermediate results (running sums, partial conclusions) sit in front of it in context instead of the whole computation being squeezed into one leap — so it errs less on multi-step tasks. You trigger it with a phrase like 'think step by step' or with few-shot examples that include worked steps. Note: modern reasoning models do this internally on their own, without being asked.
🧪 Try it yourself: take a 3-4 step arithmetic problem, run it 10 times asking for a direct answer and 10 times with 'think step by step', and score each version. Then repeat the experiment on a simple one-step question — and watch the gap disappear.
- Slapping CoT on everything, even simple extraction and classification — longer, pricier replies with no accuracy gain.
- Asking for steps + JSON in the same reply with no separation, so the steps break your parsing.
- Treating the written steps as a faithful log of the model's internal process — they're a technique that improves accuracy, not guaranteed transparency into what happened inside.
Tasks with genuine steps: arithmetic, logic, multi-condition decisions, tracing a bug through code, comparisons with trade-offs.
Simple tasks (direct classification, extraction, rephrasing) — you pay tokens and time for nothing. And with modern reasoning models you usually don't need to ask at all — try without it first.