Which tasks get better with steps, and which do not

“Think step by step” is not a general upgrade

It became a habit because it worked, and it worked because of a specific property of certain tasks. Once you know the property, you can tell in advance whether it will help, and stop pasting it onto everything.

The property is dependency. A task benefits from shown steps when the answer depends on intermediate results that must be produced before the final one can be correct.

The tasks where it helps, and why

Arithmetic and unit conversion. Every stage feeds the next. Skip the intermediate values and the last line is a guess about a calculation that never happened.

Constraint satisfaction. Scheduling across time zones, seating plans, allocating a budget across competing claims. You have to hold each constraint against each option, and doing that in writing is doing it at all.

Comparison against a list. Does this contract meet these nine requirements? Walking the list gives nine checkable claims. One paragraph gives an impression.

Multi-hop questions. “Which of our suppliers is in a country whose currency has moved more than 10% this year?” Two lookups and a join. Each hop is a place to go wrong and a place to check.

Anything where you want to audit. Even where steps do not improve accuracy, they change an opinion you must trust into work you can inspect. Frequently that is the whole reason to ask.

The tasks where it does not help

Recall. “What is the capital of Peru.” Steps add words, not accuracy.

Translation and summarising short text. The task is largely single-pass. Asking for reasoning gives you a paragraph about the translation and then the translation.

Tone and style work. Register is not produced by deduction. Reasoning about tone produces an essay about tone.

Classification with an obvious label. Most items in most classification tasks are easy. Forcing reasoning on all of them costs money and adds nothing, although a targeted version helps, below.

Creative generation. Steps narrow it. If you want twenty ideas, asking for reasoning first produces five reasoned ideas that resemble each other.

The version that beats both

Rather than always or never, make it conditional:

If the answer is obvious, give it in one line. If it requires comparing several things or several steps, show the steps first.

This works well, and it costs one sentence. On a batch of a hundred classification items you get short answers for the eighty easy ones and reasoning for the twenty that need it, which is both cheaper and more useful than either uniform policy.

Ask for the right intermediate, not just “steps”

“Think step by step” is vague, and vague instructions get vague compliance. Name the intermediate you want:

• Constraints: “List the constraints, then check each option against each constraint, then conclude.”
• Quotes: “Quote the sentence you are relying on, then interpret it.”
• Numbers: “Show the figures you used and where each came from.”
• Candidates: “List every candidate that meets the first requirement, then narrow.”

Each of these is checkable in seconds by a person who does not know the answer. That is the test of a good intermediate: not that it sounds like thinking, but that you can point at one line and say that is wrong.

What changed with reasoning models

Newer models labelled as reasoning or thinking models do this internally before answering, and on those, adding “think step by step” is usually redundant and occasionally counterproductive. That is the next lesson.

What has not changed is the second reason for asking: you wanted to see it. A model that reasons internally still gives you a final answer with the working hidden. If you need to check the UTC conversions, ask for them in the output, whatever the model does privately.

The one-line rule

Ask for steps when the answer depends on intermediate results, or when you intend to check them. Otherwise you are paying for length.

A caution about what the steps are

Everything above is about accuracy and auditability. None of it says the shown steps describe how the answer was produced.

That distinction runs through the rest of this module. Intermediate text does improve accuracy on dependent tasks: the working is doing real computational work, not narrating it. But the account of the reasoning you read afterwards is generated alongside the answer, and it can be wrong about itself. Both things are true at once, and holding both is what separates using shown reasoning well from trusting it.