Key points
- Chain of thought reaches an answer by working through the middle steps one at a time, instead of blurting it out directly.
- The key is putting those middle steps into words instead of leaving them in your head. Writing them out creates room to actually work the problem.
- Accuracy climbs on problems that need several steps. In exchange, the answer gets longer, slower, and more expensive.
- One line added to a prompt is enough, or a couple of well-worked examples can be shown instead.
- These days a reasoning model takes these steps on its own at length, while many services fold the process away or show only a summary.
The analogy
On a pool table, when the ball you're aiming for sits blocked behind another one, there's no direct shot. So a player traces a path off one cushion instead — hit here at this angle, it runs to that rail, bounces off, and lands against the target ball's side. Instead of trying to sink it in one motion, the path gets split into two or three segments.
Asking an AI a hard question works the same way. Demand the answer straight out and it's like forcing a shot at a blocked ball — it goes wide. Have it work through the middle points one at a time and each segment is short enough to land cleanly. Splitting the path to the answer into several segments and taking them one by one — that's chain of thought.
Someone who checks each rail with the cue instead of just picturing the path in their head sinks the shot more often. This method includes that same act of pulling the middle segments out where they can be seen.
In detail
Some problems can't be taken in one shot
A wide-open ball needs no cushion math at all — just hit it straight in. AI works the same way. A common greeting or a one-line fact needs no middle steps.
The trouble starts with a question tangled up in several conditions. Solve for a value and then decide something else with it, find every case that satisfies three or four conditions at once, work out how two sentences relate — questions like these. Demand a bare answer to something like this and a plausible-looking one pops out, but with nothing checked along the way, it lands right sometimes and badly wrong other times.
Split into steps, each one gets much simpler. The value settled in an earlier step becomes the starting point of the next, so the model only has to get each short stretch right.
Pulling the process out into the open is the key part
An AI repeats the same act: look at everything written so far, pick the one piece that comes next. Ask for just the answer and that's a single choice, done. Ask for the process instead, and the process itself feeds back in as input. A middle result written down earlier becomes the springboard for the next calculation.
It's the same act as checking a rail with the cue. That moment of touching it is what sets the angle. Writing out a middle sentence lets the model read back its own just-decided value and carry on from there.
That's why having it skip the process and write only the answer wipes out nearly all the benefit. The text the model actually writes out is the workspace (whether or not the user sees it on screen or it's folded away) — take the space away and there's nowhere left to work.
Two ways to ask for it
The simplest way is adding one line at the end of the question, something like "walk through it step by step." No examples needed, just an instruction — nearly free to attach.
A more reliable way is showing two or three well-worked examples first. Show the order and the level of detail used to solve a similar problem, and the model follows that same grain. This works especially well when a specific format is wanted.
Another option is spelling out the solving order directly: "first list every condition, then substitute them one by one, then sum up the answer in one line," breaking the segments apart explicitly. For work that needs review, this makes checking the result much easier.
More thinking isn't always better
Working through steps makes the answer longer. Longer means more time, and if it's called through a program, a bigger bill. Attach a process to every single question, even ones that need only a short answer, and the reader just gets tired.
The bigger concern: a wrong middle step becomes the springboard for the next one, unchanged. Misjudge the first rail and everything after, however smoothly it flows, tends to keep rolling toward the wrong spot. Some models catch their own mistake partway and correct it, but that doesn't always happen. A process that reads coherently can make a wrong answer look even more trustworthy than it should.
For anything that actually matters, it's worth skimming the process once, or solving the same problem a few times to see whether the answers drift apart.
Modern models take the steps on their own
Users used to have to ask, "think it through step by step." These days, a reasoning model takes a long process on its own once it gets a hard question. It may even work several branches, weigh them, and pick one.
How the steps get shown differs from service to service. Plenty fold them away or show a short summary instead, and some lay everything out in full. The "thinking" indicator that briefly appears before an answer lands is that time being spent.
So these days, what matters more than asking for the process is deciding how much detail to show — short for a simple question, plenty of time for a complicated one.
More precisely
Chain of thought is a prompting technique that gets a model to generate intermediate reasoning sentences before its final answer. Because it changes only the input, not the model's structure, accuracy can shift quite a bit for the exact same model depending on how the question gets phrased. When the method first became known, the effect was clear only in fairly large models. Today smaller models that were separately trained on worked-out reasoning gain from it too, but smaller ones without that training can end up with just a longer pile of off-target steps when made to write out a long process.
The analogy breaks down somewhere. In billiards, a ball hitting a rail rolls exactly by physical law, so a correctly traced path delivers the result every time. A model's middle steps come with no such guarantee. What gets written is still only a plausible next sentence, so the process on the page doesn't always match the calculation actually happening inside the model. An answer can be right with a nonsensical process, or a process can read soundly while the answer is wrong. That gap has to be kept in mind whenever the written-out process gets used as evidence that an answer is trustworthy.
Try it yourself
Common misconceptions
It's easy to think the process shown on screen is the model's real inner voice, but actually that text is also generated output, so it can diverge from what's happening internally.
It's easy to think every question benefits from step-by-step reasoning, but actually a simple question just gets a longer answer, burning extra time and cost.
It's easy to believe a coherent process means a correct answer, but actually an answer can be wrong even when the process reads smoothly. A slip in an earlier step tends to spread straight to the later ones, and the written process and the real answer can even come apart.
One-line summary
Chain of thought is like routing around a blocked ball off a cushion instead of forcing a straight shot — splitting the path to an answer into middle steps and taking them one at a time.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-17