Chain of Thought and Step-Back Prompting

Study guide

Make the AI reason before it answers

This course teaches you how to get better answers from ChatGPT, Claude and Gemini on the tasks where wrong answers cost you time: maths, logic, planning and multi step analysis. You will learn chain of thought prompting, which asks the model to reason step by step before it answers, zero-shot chain of thought, which does the same with one short instruction, step-back prompting, which asks for the general principle before the specific case, and two related methods, analogical prompting and self-ask, that make the model gather relevant knowledge before it commits to an answer. Each part explains why the method works, how to write the prompt, and what a good result looks like in a real job.

It is written for busy professionals who are not technical and who paste prompts into a chat window every day. You do not need to code or understand model internals. You will work through examples such as checking a budget calculation, drafting an onboarding plan and writing a client brief. The course also covers the change that reasoning models brought. On newer models you control depth with an effort setting and ask for visible working rather than telling the model to think step by step, because it already does that internally. You will finish knowing which method fits which task, how to verify the reasoning you receive, and when a plain prompt is the better choice.

What is a prompt framework?

A prompt is the request you type into an AI tool such as ChatGPT, Claude or Gemini. A prompt framework is a reusable structure for that request: a short checklist of the parts a good prompt contains, usually named with an acronym so it is easy to remember. Why it helps: the AI fills in whatever you leave out with generic guesses. A framework makes sure you include the parts that change the answer, such as who it is for, what you want to achieve and the format you need.

Frameworks make prompts repeatable: you fill the same parts each time, save the prompt as a template and share it with colleagues. They are not magic words: the value comes from the detail you put in each part. They help most when a request is complex or open to interpretation; a quick, simple question can stay short. There are well over sixty named frameworks, and most of them remix the same ingredients: role, task, context, audience, format, constraints and examples.

In this course you learn Chain of thought, which is a prompting technique from AI research: instead of slots to fill, it changes how the AI works through a task, and you can add it on top of any fill-in framework. It is best for: reasoning, maths, logic and multi-step analysis.

What chain of thought prompting is

Chain of thought prompting means asking the model to produce intermediate reasoning steps before it gives the final answer. The answer still comes, but you also see the path taken to reach it. The technique was introduced by Wei and colleagues at Google in 2022. Their work showed that guiding a model to reason in steps improves performance on problems that need several moves. Zero-shot chain of thought, from Kojima in 2022, showed that simply adding let us think step by step improved reasoning, with no examples or special format needed.

The gains are largest on maths, logic and multi step problems. These are tasks where one early error can lead to a wrong final answer, so visible steps help you catch mistakes before they reach your work. A practical prompt has four parts: the task, the data, the step by step instruction, and the output format. Missing any one of these weakens the result. If you leave out the data, the model invents it. If you leave out the format, you get an answer you have to reformat by hand.

Do not use step by step reasoning on simple tasks such as short rewrites or summaries. It adds length without improving quality and can make the model drift off topic. A one line request for a subject line does not need a reasoning path. The rule of thumb is simple: if you could do the task in your head in a few seconds, the model does not need steps either.

How reasoning models changed the rules

Reasoning models from Anthropic, OpenAI and Google think internally before producing an answer. This means the model already does the step by step work, so you do not need to ask for it. You control how deeply the model thinks with an effort setting or reasoning setting. Low effort is fast and cheap for simple tasks. High effort is slower but better for hard, multi step problems. Asking a reasoning model to think step by step mostly adds length, not quality, because the model was already reasoning.

Ask the model to show its working when you need to check the reasoning. This puts the logic in the answer where you can read it, question it and correct it before you use the result. Anthropic advises against over prompting thoroughness on newer models. The guidance is to choose an approach and commit to it, rather than stacking many instructions about being careful or thorough. Stacked instructions pull the model in different directions and often make the output worse.

A clean reasoning prompt has five parts: a clear role, the relevant data, an effort level, a request to show working, and a length limit. Each part has a job, and none of them repeat what the model already does. For example, you might say you are a financial analyst, paste the figures, set effort to high, ask for the working in a separate section, and cap the answer at 300 words. That prompt does more than five paragraphs of encouragement.

Step-back prompting

Step-back prompting reverses the usual order. You ask for the general principle first, then apply it to your specific case. The technique comes from Zheng and colleagues, 2023. The classic example is: What are the principles of a good onboarding plan? Now apply them to my team. The first sentence sets the standard, the second one uses it. The model names the governing idea before it produces detail, so you can judge the detail against a principle you recognise.

Step-back prompting still helps today, because it brings the right knowledge into play before the details. Reasoning models benefit from a clear frame just as much as older models did. A step-back prompt has two clear parts, the principle request and the application request. Keep them separate so the AI does not blur them together. Add your context in the second part: team size, setup, constraints. The principles stay general, but the application should be specific to you.

Check the output by mapping each detail back to a principle. If a detail does not map to anything, the AI has drifted and you should ask again. This check takes a minute and catches the most common failure, which is a list of plausible sounding items that do not follow from any stated standard. It is also a good way to explain the plan to a colleague, because you can point at the principle behind each step.

Analogical prompting and self-ask

Analogical prompting asks the model to produce similar solved problems before it attempts yours. The prompt is short: give me three examples of this done well, then do mine. The model builds patterns from those examples and applies them to your task. Self-ask has the model ask itself follow-up questions before answering. It lists what it needs to resolve, answers each question, then writes the final response. This fills gaps in the task before the model commits.

Both variants share one purpose. They make the model gather relevant knowledge before committing to an answer. Analogical prompting gathers from generated examples. Self-ask gathers by interrogating the task itself. A common failure is asking for examples and then letting the model drift. Name your audience and your constraint in the same prompt, so the model applies the right example to the right task. Without that, you get three generic examples and an answer built on the wrong one.

You can combine both. Ask for examples first, then tell the model to list the follow-up questions it still needs to resolve before applying them. This works well for tasks with hidden details, like lesson plans or client briefs, where the missing information matters as much as the visible request. Read the follow-up questions before you read the answer. They tell you what the model assumed, and you can correct those assumptions in one reply.

Showing working you can check

Separate sections keep the reasoning readable on its own. If the steps sit inside the answer, you read them already agreeing with the conclusion, which is exactly when you stop checking. Ask for the reasoning first and the answer second. That order is deliberate. Reasoning written after a conclusion tends to be shaped to fit it, which makes the check weaker. Two headings, one for working and one for the answer, are enough.

An assumption is any choice the model made that you did not state, such as scope, units, occupancy or timeframe. Tagging them turns invisible guesses into items you can confirm or correct. An unsure flag is the model telling you where its confidence is thin. Treat these as the first places to look, and ask a follow up question about each one before you rely on the result. Language models predict text, so long arithmetic is unreliable even when the sentence sounds certain.

For number work, ask for the formula, the inputs and a calculation tool, and read the method rather than the number. Checking is a review job, not a reading job. For each step ask whether the input is right, the method is right and the unit is right, then accept, correct or reject. That habit takes a few minutes and is the difference between using a figure and trusting it.

Reasoning prompts in your job

Match each method to the right task. Step-back prompting suits plans and policies because it makes the model name the governing principle before it produces detail. You can then judge the detail against a principle you recognise. Analogical prompting suits creative problems because it asks for a comparison from another field. The comparison gives you options you would not have generated from inside your own area. Show-your-working suits numbers because it exposes each figure and each step. If a step is wrong, you can point at it and ask for a correction instead of rejecting the whole answer.

Reasoning effort is a setting, not a prompt style. Raising it gives the model more room to think on a hard problem. Writing a longer prompt does not do the same thing. A long prompt full of reminders to be careful adds words, not thinking. When a hard problem comes back weak, change the setting first, then tighten the data you supplied, then adjust the wording.

Verification is still your job. Reasoning text can contain errors, so check figures against your own source and check logic against your own judgement before anything that matters goes out. The methods in this course make the model's thinking visible. They do not make it correct. The value is that you can now see where to look, ask a pointed question, and fix one step instead of starting again.

Frequently asked questions