Chain-of-Thought Prompt Generator

What this tool does

Most chain-of-thought prompts you find online are one sentence pasted between paragraphs: think step by step. The sentence did come from a pair of research papers, and it works, but it is not the whole technique. What made the difference in the studies was the shape of the prompt around that sentence: worked examples that demonstrate the reasoning, a decomposition the model has to fill in, and an answer format the result can be graded against. This tool turns your task into three of those shapes, one for each paper that tested them, so you can compare them on your own model instead of trusting a landing page that says the technique always helps. It also tells you when it will probably not help, because on short factual questions a reasoning chain mostly spends tokens.

How to use it

The prompt you need depends more on the shape of your task than on your model. Three steps get you a usable one.

1. Describe the task the way you would ask a colleague: the question itself, the numbers or constraints it depends on, and what counts as a finished answer.
2. Pick the closest task type, or leave Auto. The type controls which worked example goes into the few-shot variant and how the decomposition is phrased.
3. Take one variant to your model. Zero-shot is the cheapest way to try; few-shot suits tasks of a familiar type; structured is the one to rerun when somebody will check the steps, because it forces a second method as its own verification.

Why not just ask a chatbot

Ask a chatbot for a chain-of-thought prompt and you get a paragraph of advice, once, inside a chat you cannot keep. The technique it quotes was tested in papers whose prompts shared a structure: the task, then an explicit reasoning scaffold, then an answer format the output can be checked against. This tool builds that structure for your task in three strengths, with the citations printed next to them, and rebuilds it whenever the task changes. Nothing you type here leaves your browser; the generation is plain JavaScript, so the prompts you build stay on your machine and work offline.

When a chain of thought will not help

A reasoning chain is a hypothesis to test, not a default. On short factual questions a chain wastes tokens and can add errors by restating nearby possibilities as if they were steps. If your task can be answered with a calculator, a lookup or a small script, the verification belongs in the tool, not in prose. The other limit is honesty: the reasoning a model prints is not always the reasoning it used, and follow-up studies found the printed chain can be rationalised after the answer was already chosen. Treat the steps as material you check, not proof. If a step cites a number that is not in your task, the answer under it is decoration.

Frequently Asked Questions

What is a chain-of-thought prompt?

A prompt that asks the model to write out intermediate reasoning steps before giving the answer. In the original 2022 study, including worked example chains in the prompt lifted accuracy on grade-school arithmetic word problems from 18 percent to 58 percent.

Do I need all three variants?

No. Start with the zero-shot version: one short prompt, the cheapest to run, and often enough for a capable model. Move to few-shot when you have many tasks of the same type, and to the structured version when the steps will be inspected or the answer has to be defended.

Does this work with reasoning models like o1 or DeepSeek-R1?

Reasoning-focused models already carry out this kind of step-by-step computation internally, so a long printed chain in front of the question is mostly redundant; the structured variant is the only one that still adds something, because its verification and answer format survive the model change. For normal chat models all three variants earn their tokens.

Is my task text uploaded anywhere?

No. The page generates the prompts with plain JavaScript in your browser, and no network request is made when you click the button. Your task text never leaves the page.

Why does the structured version demand a second method?

Because recomputing the result through a second, independent method is the cheapest catch available for arithmetic slips and truncation errors, and it is a step most prompt templates leave out.

Where do the claims in the text come from?

The three prompts follow the shapes tested in the 2022 papers cited in the references, and the caution about printed reasoning being unfaithful comes from a later paper on the same technique. Each variant names its source; the numbers in the FAQ come from those papers.

References

• Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2022)
• Kojima et al., Large Language Models are Zero-Shot Reasoners (2022)
• Lanham et al., Measuring Faithfulness in Chain-of-Thought Reasoning (2023)