September 25, 2026
"Think step by step." It is the most repeated piece of prompt advice on the internet, and for good reason: on some tasks it turns a vague answer into a reliable one. On others it does nothing useful, or actively makes the output worse. The difference is not luck. It depends on whether the task has steps to think through in the first place. Here is how reasoning-style prompting works, where it earns its tokens, and where it quietly wastes them.
What Chain-of-Thought Prompting Actually Does
Output is generated one piece at a time, and each piece is influenced by everything that came before it. That single fact explains the whole technique. If you push for a conclusion immediately, the model has to compress every intermediate operation into one leap. If you ask for the working first, those intermediate values become part of the context the final answer is built on.
So the mechanism is not mysterious. It is scratch paper.
Consider a small pricing question: "A jacket costs $48 after a 25% discount. What was the original price?" Asked cold, the answer often comes back as $60 — a plausible-looking number produced by applying the discount in the wrong direction. Asked to first state the relationship ("sale price equals original times 0.75"), then solve, then round, the correct $64 appears almost every time. The steps are not decoration. They are the calculation.
The Task Types That Genuinely Benefit
The common thread is that the answer must be derived, not recalled. If the response depends on several operations chained together, reasoning prompts help. If it depends on memory or taste, they mostly add noise.
- Multi-step arithmetic and unit conversion. Currency, dosage-free measurements, time zones, percentages, pro-rated costs.
- Constraint puzzles. Scheduling five meetings around four calendars, or ordering items when two rules conflict.
- Planning with dependencies. Sequencing a product launch or a migration where step three is impossible before step two.
- Troubleshooting. Working from symptom to cause when each check depends on the result of the previous one.
- Consistent scoring. Applying the same rubric to twelve candidates or vendors without drifting between item four and item nine.
- Rule-based decisions. Anything where a stated policy determines the outcome and you want to see the policy being applied.
Equally useful: the inverse list. Summarising a document, rewriting a paragraph in a warmer tone, translating a sentence, drafting a subject line, brainstorming names — none of these improve with visible steps. You are asking for a rewrite of a single judgment call, and a fabricated procedure just sits in front of it.
How to Write a Better Reasoning Prompt
Name the steps instead of asking for "steps"
"Think step by step" leaves the procedure undefined, so the model invents one. Naming the steps removes that guesswork: "List the assumptions, compute the deduction for each year, sum them, then state the total."
Give the shape of the answer
Say how long the working should be and where it should end. Vague instructions produce rambling reasoning, and rambling is where small errors accumulate unnoticed.
Separate the working from the verdict
Ask for the conclusion on its own labelled line. You get something you can copy without rereading three paragraphs, and you can tell at a glance whether the steps support the number.
You are estimating a project budget. List each cost line with its amount, total the lines, add a 15% contingency to the subtotal, and state the final figure. Use no more than five steps. End with "Quote:" on its own line.
When It Backfires
Simple lookups. If the answer is one fact or one sentence, extra reasoning adds latency, cost, and chances to drift. A longer response is not a better one.
Subjective calls. Ask why one logo is stronger than another and you will get a tidy rationale that was written after the preference, not before it. The explanation feels rigorous and means very little.
Classification with a short label set. Order matters here. When reasoning is produced before the label, the narrative can pull the label toward whatever the reasoning drifted into. Label first, brief justification second, is usually the steadier arrangement.
Confident bad arithmetic. This is the trap people miss. A wrong answer with no working is easy to distrust. A wrong answer wrapped in four clean-looking steps reads as verified. If you cannot check the steps yourself, do not treat their presence as evidence.
Creative work. Voice and rhythm come from compression and surprise. Narrating the choices flattens both.
Tight budgets. Reasoning tokens are billed and they add seconds to every response. In a high-volume workflow, reserve them for the tasks that need them.
A Ten-Second Checklist
- Could I write the steps myself, in order? If not, you probably do not need them.
- Does the answer require three or more dependent operations? Use reasoning.
- Is there one defensible answer, or a matter of preference? Be cautious with the second.
- Will I actually read the steps? If not, ask for a one-line justification instead.
- Is speed or cost a hard constraint? Then keep it short or skip it.
Putting It Into Practice
Run the same prompt twice on a task you care about — once plain, once with named steps — and compare. Beyond a few tries, the pattern becomes obvious: arithmetic, planning and rule-following sharpen; retrieval, style and taste do not.
Keep your working prompts in a short saved list so you are not rebuilding them each time, and treat the steps as something to inspect rather than admire. The point of asking for the working is that it makes an answer checkable. When you cannot check it, ask for less.
Photo: Walls.io / Pexels
