Chain-of-Thought: step-by-step AI — Prompts

What for: raise accuracy on reasoning tasks (logic, maths, analysis) — the model thinks it through first and only then answers.

When to use it

Any task with several reasoning steps: maths, logic, calculations, unpacking conditions, "work it out and explain". The role is analyst/solver. Result: the model doesn't jump to an answer, it talks through the reasoning — and gets things wrong noticeably less often.

The prompt (copy and paste)

Solve the task below, reasoning STEP BY STEP.
TASK: "<PASTE THE TASK>"

Rules:
1. First break the task into steps out loud and talk through the reasoning for each.
2. At every step state explicitly which data you're using and what interim conclusion you reach.
3. If data is missing, say exactly what's missing and don't make it up.
4. Only AFTER the reasoning give the final answer, separated by a line reading "Bottom line:".

The technique: putting the reasoning into text (CoT) cuts errors on multi-step tasks — the model doesn't jump straight to the answer.

Filled-in example

Instead of <PASTE THE TASK>: "A shop bought 200 T-shirts at 350 ₽, sold 170 at 700 ₽, and wrote off the rest. What's the net profit and the margin?"

Expected AI response (single pass): Step 1 — purchase 200 x 350 = 70,000 ₽. Step 2 — revenue 170 x 700 = 119,000 ₽. Step 3 — 30 units written off; in purchase terms that's already in the costs. Step 4 — profit 119,000 − 70,000 = 49,000 ₽. Step 5 — margin 49,000 / 70,000 ≈ 70%. Bottom line: profit 49,000 ₽, margin ~70%.

Variations

Self-consistency. Add "Solve the task three independent ways and compare the answers; if they diverge, find the mistake" — for critical calculations.
Hide the reasoning. If you only want the result: "Think step by step to yourself, but show me only the Bottom line block".
Auditing someone else's solution. "Here's a solution and its answer. Check every step and point out where the first error appears."

Pro tips

On reasoning models (the o-series, Claude with extended thinking) CoT is built in — a separate instruction is barely needed; the effect is more pronounced on regular models.
Don't use it for simple factual questions — step-by-step only bloats the answer.
Common mistake: trusting the "Bottom line" without reading the steps. The whole value of CoT is that the error is visible in the reasoning — so read the reasoning.