Chain of Thought Prompting: 5 Examples That Ship (2026)
Five paste-ready chain-of-thought prompting examples for 2026, each with its failure mode, plus when reasoning out loud makes the answer worse.
On this page
Quick answer
Chain of thought prompting is asking the model to reason step by step before it answers, so the hard part happens on paper instead of in one guess. In 2026 it still lifts accuracy on multi-step math, logic, and extraction, and still does nothing (or worse) on trivial tasks and on reasoning models that already think internally. Below are five paste-ready chain of thought prompting examples, each with the failure mode nobody warns you about, plus when to skip it.
OpenAI's prompting guide point the same way: give the model room to reason before the answer. The technique traces back to the 2022 Wei et al. paper. Everything below is the practical version.
Recipe 1: Zero-shot CoT
The cheapest reasoning boost. One line.
Q: A shop sells pens at 3 for $2. I buy 12. How much do I pay?
Let's think step by step.
Why it works. The trigger phrase makes the model lay out intermediate steps. Intermediate steps catch arithmetic and unit errors a one-shot guess skips.
Fails when: the task is trivial. On a lookup or a yes/no, the model invents steps to justify an answer it already had, and sometimes talks itself out of the right one.
Recipe 2: Few-shot CoT
Show the reasoning, do not just ask for it.
Q: Roses are red. All red things are warm. Are roses warm?
A: Roses are red. Red things are warm. So roses are warm. Answer: yes.
Q: Cats are mammals. All mammals breathe air. Do cats breathe air?
A: Cats are mammals. Mammals breathe air. So cats breathe air. Answer: yes.
Q: Sharks are fish. All fish live in water. Do sharks live in water?
A:
Why it works. Two worked examples teach the shape of the reasoning, not just the format of the answer. The model copies the pattern on the new input.
Fails when: your exemplars share a shortcut the real input does not. The model pattern-matches the shortcut instead of reasoning. Vary your examples or it learns the wrong path.
Recipe 3: Structured CoT
Force the reasoning into named fields. Then a clean answer.
Extract the reasoning, then answer. Use exactly these fields:
FACTS: (list only what the text states)
INFERENCE: (what follows from the facts)
ANSWER: (one line, final)
Text: {input}
Why it works. Named fields stop the model from blending reasoning and conclusion into one hedge. You can read FACTS to see if it hallucinated before you trust ANSWER.
Fails when: the model fills the fields as theater. It writes plausible FACTS, then ignores them in ANSWER. Always check that ANSWER actually follows from INFERENCE.
Recipe 4: Self-consistency
Sample the same CoT prompt several times. Take the majority answer.
Solve step by step. End with "Answer: X".
(run this prompt 5 times at temperature 0.7, keep the most common final answer)
Why it works. Independent reasoning paths disagree on the wrong answers and agree on the right one. The vote filters one-off slips.
Fails when: the model is confidently wrong the same way every time. A shared blind spot survives the vote. Self-consistency fixes noise, not bias.
Recipe 5: CoT then extract
Reason in a scratchpad. Return only the final value.
Think through this in a .
After the scratchpad, output ONLY the final answer as JSON: {"result": ...}
Task: {input}
Why it works. You get the accuracy of reasoning and a clean value to parse. The messy thinking stays out of your response payload.
Fails when: you parse the scratchpad instead of the final line. Grab the JSON after the reasoning, never the reasoning itself, or you ship the model's out-loud doubt to your user.
When to skip chain of thought
CoT is not free. Skip it here.
Reasoning models already think. On o-series and extended-thinking models, the reasoning runs internally. Bolting "think step by step" on top burns tokens and can degrade the built-in process.
Hot paths and latency budgets. CoT multiplies output tokens. On a request that fires on every keystroke, that cost compounds fast.
Simple tasks. Classification, formatting, and lookups do not need a monologue. Reasoning out loud gives the model more room to overthink a call it had right.
Chain of thought is a tool for multi-step problems. Use it where the steps are real. Everywhere else, ask for the answer and move on. For splitting reasoning across separate calls, see the prompt chaining recipes; for teaching the shape with examples, the few-shot recipes.
Cost to test: $0.02 to run all five examples once on a small 2026 model, most of it the self-consistency pass.
Written by
Roo IyerRoo Iyer writes terse, contrarian prompt recipes for production builders. Opinionated about cost math.
FAQ
What is an example of chain of thought prompting?
The simplest example is zero-shot CoT: append 'Let's think step by step' to your question. The model then writes out its intermediate reasoning before the final answer, which catches arithmetic and logic slips a one-shot guess misses.
What is the difference between chain of thought and few-shot prompting?
Chain of thought asks the model to reason step by step inside one answer. Few-shot prompting shows the model worked examples of the task. They combine: few-shot CoT gives examples that include the reasoning, teaching the shape of the thinking, not just the format of the answer.
Does chain of thought prompting work on reasoning models?
Mostly no. Reasoning models like the o-series and extended-thinking models already reason internally before answering. Adding 'think step by step' on top is redundant, burns extra tokens, and can degrade the built-in process. Save CoT prompting for standard models.
When should you not use chain of thought prompting?
Skip CoT on trivial tasks (classification, formatting, lookups), on latency-sensitive hot paths where extra output tokens compound, and on reasoning models that already think internally. On simple calls, reasoning out loud just gives the model room to overthink a correct answer.
What is zero-shot chain of thought prompting?
Zero-shot CoT is triggering step-by-step reasoning with a single instruction, classically 'Let's think step by step', without providing any worked examples. It is the cheapest reasoning boost and a good default before you reach for few-shot CoT.
Does chain of thought prompting make models more accurate?
On multi-step reasoning, math, and structured extraction, yes: the 2022 Wei et al. paper showed clear gains. On simple one-step tasks it adds no benefit and can slightly hurt by letting the model talk itself out of a right answer. Match the technique to the task.
Related recipes
Prompt Chaining: 5 Recipes That Ship (2026)
Five paste-ready prompt chaining recipes for 2026: extract-then-reason, draft-critique-rewrite, route, map-reduce, and generate-then-gate, each with its failure mode.
Few-Shot Prompting Recipes That Ship (2026)
Five copy-paste few-shot prompting recipes for July 2026: lock output format, tone, labels, and edge cases. Each with the prompt and the exact case where it breaks.
Meta Prompting: 5 Recipes That Ship (2026)
Five paste-ready meta prompting recipes: make the model write, critique, rubric, and template your prompts, each with its failure mode. Plus the honest part: when meta prompting just wastes tokens.