Kwabena owns a landscaping company in Atlanta with eight employees. He had been using ChatGPT for a few months, mostly for writing, and it worked fine for that. Then a big commercial client asked for a proposal that covered plant selection for their climate zone, an estimate for installation labor and materials, a maintenance schedule for the first year, and a payment terms section. Kwabena typed "write a proposal for a commercial landscaping client" and got back something generic and useless. He tried again with more detail and got slightly less generic and slightly less useless. He was about to give up on AI for anything complicated when a supplier mentioned the phrase "chain-of-thought prompting." That phrase, and the technique behind it, changed everything.

Why AI Struggles with Complex Tasks

When you give a person a complex request, they break it into steps without being told to. You ask a contractor to "plan the kitchen renovation" and they automatically think: budget first, then layout, then permits, then timeline, then materials, then subcontractors. Nobody instructed them to sequence the work that way. Years on the job taught them that a big question is really several smaller questions asked in a sensible order, and they run that sequence silently before saying a word back to you.

AI does not do this automatically. It responds to what you ask. If you ask one vague question, it gives one vague answer. But if you guide it to reason through a problem step by step, which is what chain-of-thought prompting means, it produces dramatically better results on complex tasks. You are not changing what the AI can do. You are changing how you ask it to think. The capability was already there. What was missing was the instruction to use it in order.

The analogy is the difference between handing a new employee a task with no instructions and walking them through your thought process first. Same employee, same capability, very different output. Kwabena's first attempt failed for exactly this reason. He asked for the finished artifact, the proposal, without ever describing the reasoning that produces a good one. The AI obliged by producing the shape of a proposal with none of the thinking underneath it.

The Core Technique

Chain-of-thought prompting has one essential element: you explicitly ask the AI to think through the problem in steps before giving you an answer. Everything else is variation on that idea. There are three practical ways to do it, and they differ mainly in how much of the reasoning you supply yourself. The first is the lightest touch, the second gives you checkpoints, and the third puts your own professional judgment into the instruction.

Method 1: The "think step by step" instruction

Simply add "Think through this step by step before answering" to any complex prompt. This single addition often improves output significantly on multi-part problems. The AI surfaces its reasoning before landing on conclusions, which lets it catch contradictions and produce more coherent answers. It costs you nothing but a sentence, so it is the first thing to try whenever a request has more than one moving part.

Instead of "Write a proposal for a commercial landscaping client," Kwabena wrote: "Think through this step by step before writing. I need a commercial landscaping proposal for an Atlanta office park. The client wants: plant selection appropriate to zone 7b, labor and materials estimate for a 2-acre installation, a twelve-month maintenance schedule, and payment terms. Walk through each section logically, then write the proposal." The request is the same. The instruction to sequence it is what changed.

What this method cannot give you is a checkpoint. The AI reasons and answers in one continuous pass, so if it adopts a wrong assumption early, everything after it inherits the error and arrives looking just as considered as the parts that are correct. You can read the reasoning afterwards and catch the problem, but by then the whole draft has been built on it. That limitation is the reason the second method exists, and it is why a single well-reasoned pass is a starting point rather than a finish line on high-stakes work.

Method 2: Staged prompting

Break the task into separate prompts and run them in sequence. Each prompt builds on the output of the previous one. This works better than a single long prompt when each stage requires genuinely different thinking, and it gives you something the single prompt cannot: a place to stop, read, and correct before the error compounds. Plant selection is a horticulture question. Labor estimating is an arithmetic question. They do not benefit from being answered in the same breath.

For Kwabena's proposal, staged prompting looks like this:

• Prompt 1: "List the plant species most suitable for a commercial property in USDA zone 7b in Atlanta. Prioritize low maintenance, drought tolerance after establishment, and visual appeal year-round. Give me twelve options with a brief note on each."
• Prompt 2 (after reviewing the plant list): "Based on these plants [paste the list], estimate materials cost and labor hours for installation on a 2-acre commercial property. Assume I am providing all plants and materials. Use typical Atlanta labor rates of $35 per hour. Show your calculation."
• Prompt 3: "Write a twelve-month maintenance schedule for the plantings we selected, covering watering, fertilizing, pruning, and seasonal preparation."
• Prompt 4: "Draft a payment terms section for a commercial landscaping contract: 30% deposit at signing, 40% at installation completion, 30% at thirty days. Include a late payment clause."
• Prompt 5: "Combine the outputs from the previous steps into a professional proposal. Use a business letter format. Keep total length under two pages."

Notice the shape of that sequence. The first four prompts each produce one component, and only the fifth assembles. Notice also the bracketed instruction in Prompt 2: [paste the list] is not decoration, it is the handoff. Staged prompting only works if you actually carry the reviewed output of one stage into the next, rather than trusting the AI to remember it. This takes longer than a single prompt, but the output is dramatically more accurate and specific, because each stage can be verified and corrected before the next one builds on it.

Method 3: Providing the reasoning framework yourself

You tell the AI exactly what to consider, in what order, before generating the answer. This is the most controlled version and works best when you know the domain well. It is also the version that scales, because a framework you write once can be pasted into every instance of the same decision for as long as your business runs it that way.

Kwabena's version reads: "When I give you a pricing decision, I want you to: 1) identify my direct costs for materials and labor, 2) add my standard 35% markup, 3) check whether the result is competitive with typical Atlanta market rates, 4) flag any items where the cost seems unusually high. Here is the job I need priced: [details]."

You are giving the AI your own decision process. It applies that process consistently every time, without you having to rethink it. The numbered steps matter more than the wording: they force the work into an order, and the fourth step, the flag, is what turns the AI from a calculator into something closer to a second reader. Note that this method inherits your expertise and also your blind spots. If your pricing logic has a gap, the AI will reproduce that gap flawlessly on every job.

Where Chain-of-Thought Makes the Biggest Difference

This technique is not necessary for simple tasks. "Write a thank-you email" does not benefit from chain-of-thought, and adding the instruction anyway just makes the AI narrate obvious reasoning at you before producing the four sentences you wanted. Reserve it for the requests where the answer genuinely depends on a chain of intermediate judgments. It makes a substantial difference for:

• Estimates and quotes that involve multiple cost components
• Decision-making prompts where you want the AI to weigh tradeoffs
• Documents that combine multiple distinct sections (proposals, reports, contracts)
• Any analysis where the quality of the conclusion depends on the quality of the steps leading to it
• Troubleshooting problems where the cause is not obvious

Those five categories share a family resemblance. In each of them, a wrong intermediate step quietly poisons everything downstream, and the finished output looks just as confident either way. That is precisely the situation where you want the intermediate steps visible.

The cost of skipping the technique on those tasks is rarely a visibly bad answer. It is a plausible one. Kwabena's first attempt did not come back obviously broken; it came back generic, which is much harder to argue with and much easier to send. An estimate assembled without visible steps still totals correctly, a proposal without sequenced sections still reads like a proposal, and a tradeoff analysis without stated weighing still reaches a conclusion. What you lose is any way of telling whether the conclusion was earned, which matters most on exactly the documents a client will read closely.

Choosing Between the Three Methods

The three are not mutually exclusive. Kwabena's staged sequence includes a "show your calculation" instruction inside Prompt 2, which is Method 1 operating inside Method 2. Start at the top of that table and move down only when the lighter method leaves you correcting the same thing twice.

Checking the Reasoning

One underused benefit of chain-of-thought prompting: when the AI shows its reasoning, you can catch errors before they end up in your final output. If Kwabena asks the AI to estimate labor hours and show its calculation, he can see whether the assumption of eight hours per thousand square feet matches his actual experience, and correct it before the estimate goes to the client. The assumption is the thing to inspect. The multiplication is rarely where these go wrong.

This is also the point at which chain-of-thought stops being a prompting trick and becomes a working habit. Kwabena's advantage over the AI is not that he writes better prompts; it is that he has installed plantings in Atlanta and knows what a crew actually gets through in a day. A visible reasoning trace is the only place that knowledge can be applied, because it is the only part of the output where his experience and the AI's assumption are stated in comparable terms. Hide the reasoning and his expertise has nothing to attach to.

This is not just error-catching. It is a form of professional review. You are not accepting the AI's output blindly; you are reviewing its logic, as you would review the work of a junior staff member. The fact that the reasoning is visible makes that review possible. An answer with no working shown gives you nothing to disagree with, which is not the same thing as being right.

Build the review into the habit rather than leaving it to good intentions. When a stage produces numbers, read the assumptions line before you read the total. When a stage produces a list, check whether anything you would have included is missing, since omissions are invisible in a confident answer. When a stage produces a recommendation, ask what it weighed and what it discarded. Then, and only then, paste it into the next prompt.

Anti-Patterns to Avoid

• Bolting "think step by step" onto a genuinely simple request. The instruction has a cost in verbosity and reading time. A thank-you email does not need a reasoning trace; save the technique for requests with real intermediate judgments.
• Running all five stages back to back without reading any of them. Staged prompting earns its extra minutes only through the review between stages. Pasting stage one straight into stage two unread produces a longer version of the single-prompt failure, and it costs more.
• Treating a visible calculation as a verified calculation. Shown work invites review; it does not perform it. Kwabena's eight-hours-per-thousand-square-feet assumption is exactly the kind of plausible-looking input that survives unchallenged into a client-facing estimate.
• Writing a reasoning framework for a domain you do not know. Method 3 transmits your judgment to the AI. If the judgment is guesswork, you have built a machine for applying guesswork consistently, which is worse than applying it once.
• Referencing earlier output instead of pasting it. "Use the plant list from before" invites the AI to reconstruct rather than reuse. Paste the reviewed text into the bracketed placeholder every time.
• Letting the assembly prompt rewrite the components. A final "combine these" prompt should join and format what you approved, not regenerate it. If the combined draft contains figures you never saw in an earlier stage, the chain broke somewhere.

Practice Prompts

Run these against a real job of your own, not a hypothetical one. The technique only teaches you anything when you can tell whether the answer is right.

• Baseline comparison. "Write a proposal for [your most recent complex client request]." Save the output. Then run the same request prefixed with "Think through this step by step before writing," and put the two side by side.
• Framework extraction. "I am going to describe a decision I make regularly in my business. Ask me questions until you can write it as a numbered list of steps I follow, in order. Here is the decision: [describe it]."
• Assumption surfacing. "Estimate labor hours for [your job type and size]. Before the total, list every assumption you are making about rates, productivity, and scope, one per line. Show your calculation."
• Stage split. "Here is a document I need: [describe it]. Do not write it. Instead, break it into the smallest number of stages where each stage needs genuinely different thinking, and tell me what I should check after each one."
• Assembly only. Take four approved component outputs and run: "Combine the following into a single professional document. Do not add new facts, figures, or sections. Use a business letter format. Keep total length under two pages."

Reflection Questions

• Which of your recurring requests have you been asking as one vague question because breaking them apart felt like extra work?
• When the AI last produced an estimate for you, did you check the assumptions or only the total?
• Is there a decision you make often enough that writing your reasoning framework once would pay for itself? What are its steps, in order?
• If you handed your framework to a new employee, would they reach your answer? If not, what did you leave out?
• Where in your current workflow would an early error go unnoticed until a client saw it?

Glossary

• Chain-of-thought prompting: Explicitly asking the AI to reason through a problem in steps before it produces an answer, rather than requesting the finished answer directly.
• Staged prompting: Splitting a complex task into a sequence of separate prompts, where each one builds on the reviewed output of the previous one.
• Reasoning framework: Your own decision process, written out as ordered numbered steps and given to the AI so it applies the same logic to every instance of a recurring decision.
• Intermediate output: The result of one stage in a sequence, produced to be checked and carried forward rather than delivered to a client.
• Assembly prompt: The final prompt in a staged sequence, whose job is to join and format approved components without generating new content.
• Bracketed placeholder: A marker such as [paste the list] or [details] inside a reusable prompt, showing where you substitute the specific content for this run.

Related Lessons

• The Anatomy of an Effective Prompt covers the components of a single well-built prompt, which is the unit each stage of a chain is made from.
• Advanced Prompt Structures and Chaining extends sequencing beyond a single document into longer connected structures.
• Multi-Step AI Workflow Architecture takes the same staging idea and treats it as an operational workflow rather than a one-off exercise.
• Common Prompting Mistakes and How to Fix Them catalogues the failure modes that show up before a prompt is complex enough to need chaining.
• Building Your Personal Prompt Library is where a framework you have tested should be stored so you are not rewriting it each time.

Closing

Kwabena did not get a better AI tool. He got the same tool to work through the proposal the way he would have worked through it himself, one component at a time, checking each before moving to the next. The commercial proposal that had defeated him twice came together because the request finally matched the shape of the work. That is the whole idea: complex output requires sequenced input, and the sequence is yours to supply.

Pick one request you have already given up on and rebuild it as stages this week. You will spend more minutes on it than you did on the vague version, and you will spend far fewer than you did rewriting what the vague version returned.

Key Takeaways

• AI responds to how you ask, not just what you ask. Vague complex prompts produce vague complex answers. Structured prompts produce structured useful answers.
• Add "think step by step" to any complex prompt. This single instruction often improves output quality significantly on multi-part problems.
• Use staged prompting for tasks with distinct sections. Run one prompt per section, review each before moving on, then combine in a final prompt.
• Give the AI your reasoning framework for decisions you make regularly. It will apply your logic consistently, faster than you can each time.
• Visible reasoning lets you catch errors before they reach clients. When the AI shows its work, you can verify it, just as you would with a junior employee's calculation.
• Chain-of-thought is most valuable for estimates, multi-section documents, and analysis tasks. Simple writing tasks do not need it.
• Staged prompting takes longer but produces better results on complex jobs. The extra prompts are an investment in output quality, not wasted time.

Frequently Asked Questions

• Does chain-of-thought prompting require a specific AI tool? No. It is a way of writing the request, not a feature you switch on. The instruction to reason in steps, the decision to split a task into stages, and the numbered framework all live entirely in the text you type.
• How do I know whether a task is complex enough to need this? Ask whether the answer depends on intermediate judgments that could each be wrong on their own. Estimates with several cost components, documents combining distinct sections, tradeoff decisions, and troubleshooting all qualify. A single short piece of writing does not.
• Is staged prompting worth the extra time? On complex jobs, yes. The extra prompts buy you checkpoints, and a corrected assumption at stage two costs a minute while the same error discovered in a client-facing proposal costs considerably more than that.
• What if the AI's reasoning looks right but the conclusion is wrong? Read the assumptions rather than the arithmetic. Reasoning traces usually fail at an input that sounds plausible, such as a productivity rate or a market comparison, not at the step that combines them. Correct the assumption and rerun that stage.
• Can I reuse a reasoning framework across different jobs? That is the point of writing one. A framework describes the decision, not the instance, so the same numbered steps apply to every job of that type. Revisit it when your costs, markup, or market change.