intellimetrics Learning
Mastery Module 12 · The Prompting Playbook

Let the Model Think

Last reviewed

Intermediate

What you'll learn

~14 min
  • Prompt for step-by-step reasoning on tasks where logic determines the answer
  • Separate the reasoning from the final answer so output stays clean
  • Tell the difference between a thinking prompt and a thinking-effort setting

A model that answers immediately commits to its first instinct. For anything with logic in it — arithmetic, multi-step analysis, a decision with competing factors — that first instinct is often wrong, and once it is written, the rest of the response rationalizes it.

The fix is one line:

Think step by step before giving your answer.

Anthropic’s workshop calls this precognition — letting the model do its reasoning out loud first, so the final answer is the conclusion of a worked-through process rather than a guess it then defends. Asking for the reasoning measurably improves accuracy on tasks that have a right answer.

Why writing the reasoning helps

The model generates one token at a time, each conditioned on what came before. If the answer comes first, there is no worked reasoning for it to rest on. If the reasoning comes first, the answer is generated after the relevant facts and intermediate steps are already on the page — so it has something correct to build on.

In practice: “ask, then answer” gives you a snap judgment. “Reason, then answer” gives you a considered one. The cost is a few extra seconds and some tokens; the payoff is fewer confidently wrong answers.

Order matters: reasoning before the answer

This only works if the reasoning genuinely comes before the conclusion. If you ask for “the answer, then your reasoning,” the model picks the answer first and the reasoning becomes after-the-fact justification — which defeats the point.

Weak (answer first):

What's the total? Give the number, then explain how you got it.

Strong (reasoning first):

Work out the total step by step — list each line item and a running
subtotal as you go — and give the final number only at the end.

Keep the reasoning out of the final output

Step-by-step reasoning is useful to generate but often noise to read. You usually want the model to think, then hand you only the conclusion. Give the reasoning a place to live and tell the model where the answer goes:

First, reason through this in a <scratchpad>. Then give your final
recommendation after the scratchpad, in 2–3 sentences.

Now you get the benefit of worked reasoning, and your output is clean — you can read the recommendation and ignore (or audit) the scratchpad. This pairs directly with the tags you’ll learn in the next lesson.

ℹThinking prompt vs. thinking-effort setting — not the same thing

Many tools now expose a thinking-effort control — a flag, a slider, or “extended thinking” mode (Module 10 mentioned this). That setting decides how much internal computation the model spends, often in a hidden reasoning channel you never see.

Asking the model to “think step by step” in the prompt is different: it shapes the visible response so the reasoning appears in the output where you can read, check, and reuse it.

They stack. On a hard problem, raise the effort setting and ask for visible step-by-step reasoning — the setting buys depth, the prompt makes that depth inspectable. Use the visible version whenever you need to see and verify the reasoning, or when your tool has no effort control.

When to use it — and when to skip it

Use step-by-step reasoning when the answer depends on a chain of logic:

  • Math, totals, unit conversions, anything quantitative
  • Decisions with trade-offs (“which vendor, given these constraints?”)
  • Multi-step analysis (“does this code have a race condition?”)
  • Anything where you’ve seen the model jump to a wrong conclusion

Skip it when the task is recall or simple transformation — “what’s the capital of France,” “reformat this list as bullets.” There’s no reasoning chain to walk, so asking for one just adds latency and clutter.

🧬In Your Field: Biotechclick to expand

Make the calculation auditable. For dilution series, molarity, or coverage math, “show each step and the formula you used before the final value” turns a black-box number into something a second scientist can check. In the lab, the worked steps are not optional — they’re the record.

📊In Your Field: MIS / Businessclick to expand

Reason before the verdict on judgment calls. For a build-vs-buy or vendor-selection decision, “list the factors for and against each option, weigh them, then give the recommendation” produces a defensible call instead of a gut reaction you can’t interrogate later.

KNOWLEDGE CHECK

You want better accuracy on a multi-step budget calculation AND a clean final number. What's the best prompt structure?

Key takeaways

  • Reason, then answer. On anything with logic, telling the model to work through its reasoning before committing measurably improves accuracy.
  • Order is the whole trick. Reasoning must come before the conclusion, or it becomes after-the-fact justification.
  • Park the reasoning in a scratchpad so you get the accuracy benefit without cluttering the output — and so you can audit it.
  • A thinking prompt ≠ a thinking-effort setting. The setting buys hidden depth; the prompt makes reasoning visible. Stack them on hard problems.
  • Skip it for recall and simple reformatting — there’s no logic chain to walk.

Next: Structure and Format

Search lessons