Give it room to think

Ask someone "what's 23 plus 48 minus 39?" and demand the answer immediately — no pause allowed. They'll probably blurt something plausible and wrong. Given three seconds to work through it, they'll nail it.

Language models are the same, for a very mechanical reason. A model spends a roughly fixed amount of compute per token it generates — that's the forward pass you may have met in the inference course. If the very first token of its reply has to be the answer, the entire problem must be solved in that single step. But let it write out intermediate steps, and each small step only needs a little reasoning — with every previous step sitting right there in the context to build on.

Scroll to see the same question answered both ways.

Thinking out loud is the computation

This technique is called chain-of-thought prompting, and the phrase "let's think step by step" is its famous magic-sounding form. But you now know it isn't magic: tokens are compute, so asking for visible steps literally buys the model more computation — and lets it check each move against the ones before it.

It pays off on anything multi-step:

• arithmetic and unit conversions
• logic puzzles and scheduling constraints
• debugging ("walk through what this function does with input X")
• decisions with tradeoffs ("list pros and cons for each option, then recommend")

A useful variant: ask for the reasoning first and the answer last. If the conclusion comes first, everything after it is just rationalization.

A café starts with 23 muffins, bakes 4 trays of 12,
and sells 39. How many are left?

Work through this step by step, showing each
calculation. Only then give the final answer.

Don't models do this by themselves now?

Increasingly, yes. Modern "reasoning" models are trained to generate a hidden chain of thought before answering — the trick worked so well it got baked in. But the principle still shapes your prompts: hard problems deserve room, "answer in one word" quietly forbids thinking, and asking the model to show its work remains your best tool for catching where a wrong answer went off the rails.

So far each technique has been a single move. Next, we assemble the whole playbook into one well-built prompt.