Chain-of-Thought & Reasoning, Explained From Zero

One tiny addition to a prompt — four words — can dramatically improve a model's accuracy on anything involving logic. This guide starts from nothing, shows why the trick works, and gives you the judgment for when it's worth the extra time and when it's just waffle.

1 · The model that blurts the answer

Ask a model a question that needs a few steps of reasoning — a bit of math, some logic, a multi-part problem — and by default it does something risky: it blurts an answer immediately, in one shot, without visibly working through it. Like a person forced to answer a tricky puzzle instantly, it gets these wrong far more often than it should. The knowledge is in there; the problem is that jumping straight to the answer skips the working. The fix for this whole class of errors is beautifully simple, and it's one you already use when a problem is hard: slow down and think it through step by step.

2 · Chain-of-thought: think step by step

Chain-of-thought prompting means asking the model to reason out loud, step by step, before its final answer. The magic phrase is almost comical: add “let's think step by step” or “show your reasoning.” Instead of leaping to a conclusion, the model writes out the intermediate steps — and writing each step makes the next more likely to be correct, because each becomes solid ground rather than a leap. For “twelve apples, eat four, buy five,” it writes: start with twelve, eat four leaves eight, buy five makes thirteen. Same model, same question, but by laying down stepping stones, accuracy on reasoning tasks jumps. You didn't make the model smarter; you gave it room to use the intelligence it has.

3 · Why it actually works

Why does merely writing steps help so much? A model generates one piece of text at a time, and each new piece is shaped by all the text before it. A model has no private scratchpad, no inner mind where it quietly works things out — its “thinking” can only happen on the page, in the words it writes. So when you let it write the intermediate steps, those steps become the context that guides the next step. Force it to skip straight to the answer and you've denied it the very medium it thinks in. Chain-of-thought works because it lets the model reason where it actually can — out loud.

4 · When to use it — and not

Chain-of-thought isn't free, so use it where it pays. Reach for it whenever a task involves several steps: arithmetic and math, logic, planning, debugging, comparing options against criteria — anywhere a wrong intermediate step derails the answer. There, the accuracy gain is huge. But skip it for simple things — a direct fact, a quick lookup, a translation, casual chat. Forcing “think step by step” onto “what's the capital of France” just produces waffle, wastes time, and costs more tokens. The judgment is the one you'd apply to yourself: hard, multi-step problem, slow down; easy, familiar question, just answer.

5 · A worked example: the trap question

The classic example shows the whole effect. A bat and ball cost $1.10 total; the bat costs $1 more than the ball; how much is the ball? Almost everyone's gut screams “ten cents” — and a model answering instantly usually blurts the same, because it feels right. But it's wrong. Force the working: let the ball be x, so the bat is x + $1; together $1.10, so 2x + 1 = 1.10, meaning 2x = $0.10, so x = $0.05. The ball is five cents. Same model, same question — the only change was laying down steps instead of trusting the snap answer. That gap between the fast wrong answer and the worked right one is exactly what chain-of-thought closes.

6 · Reasoning models: built-in thinking

An important update: many newer models are reasoning models that do this automatically. Behind the scenes, they take time to think step by step before answering, without being asked — the technique has been baked into the model. So you may not need to type “think step by step” with them. But two things stay true: understanding chain-of-thought explains why these models are better at hard problems and why they take longer and cost more (they're literally doing more thinking); and your job doesn't vanish — you still have to state the problem clearly and give the needed context, because even a great reasoner can't reason well about a vague question.

7 · You can read the reasoning

A valuable bonus: when the model shows its steps, you can check them. If the answer looks off, read the reasoning and you can often pinpoint exactly where it went wrong — a bad assumption in step two, a dropped detail in step four. That makes answers far more trustworthy and debuggable than a bare number with no explanation. It also protects you from the most dangerous thing in AI: a confident wrong answer. So even when you don't strictly need the reasoning, asking for it on important questions is cheap insurance — visible reasoning is checkable reasoning. Treat the steps like a colleague's working: usually right, but worth a glance before you rely on the conclusion.

8 · A simple test you can run this week

9 · Glossary — every term, spelled out

Chain-of-thought

Prompting a model to write out its reasoning step by step before giving a final answer.

Reasoning model

A model that automatically thinks step by step before answering, without being asked — slower and costlier, but better at hard problems.

Thinking on the page

The fact that a model's reasoning can only happen in the words it writes; it has no private scratchpad.

Intermediate step

A single reasoning step written out, which becomes solid ground for the next.

Confident wrong answer

A plausible-sounding but incorrect reply given with no working — the risk chain-of-thought helps you catch.

Verifiable reasoning

Visible steps you can check, making an answer trustworthy and debuggable.