Save Extended Thinking for Genuinely Hard Problems

A judgement call about when to turn on extended thinking — the mode where Claude reasons at length before answering. Reach for it on problems that genuinely need multi-step reasoning and where a wrong answer is costly; skip it on simple or easily-checked tasks where it just adds latency.

Extended thinking trades time for depth: it lets the model work through intermediate steps rather than answering from the first plausible path. That's a real gain on hard maths, tricky debugging, planning, and thorny trade-offs — and pure overhead on 'rewrite this politely', where the extra deliberation changes nothing but the wait. Knowing which is which gets you the depth when it matters without paying for it when it doesn't.

Multi-constraint problems, subtle bugs, architecture and planning, hard analysis, anything where the answer has to be right and checking it is expensive.

Simple lookups, formatting, short rewrites, and high-volume calls where latency compounds. If a wrong answer is cheap to catch and redo, thinking rarely earns its cost.

This problem needs careful reasoning: <the hard problem, with constraints>. Think it through step by step before answering — consider more than one approach, check your reasoning against the constraints, then give me the conclusion and the key steps that got you there.

Scheduling shifts under overlapping availability, seniority, and labour-law constraints, extended thinking works through the conflicts and produces a valid schedule with its reasoning — whereas a fast answer confidently returns a roster that violates two rules.

Ask Claude to expose the decisive steps of its reasoning, not the whole trace — the two or three inferences the conclusion actually hinges on — so you can audit the logic without wading through everything, and catch a wrong turn where it happened.

• Turning on thinking for trivial tasks and paying latency for no quality gain.
• Skipping it on genuinely hard problems and accepting a confident first-path answer.
• Reading the conclusion but never checking the decisive steps on high-stakes work.