2025-12-14 – Extended Thinking in Claude: A Reasoning Guide

✅ Extended Thinking — When Deeper Reasoning Pays Off

Extended thinking allows Claude to reason through a problem step by step before producing its final answer — an internal scratchpad that the model uses to explore approaches, spot errors in its own reasoning, and arrive at more reliable conclusions on hard problems. The feature is available on Claude 3.7 Sonnet and is activated by setting "thinking": {"type": "enabled", "budget_tokens": N} in your API call. The budget_tokens parameter controls how many tokens Claude is allowed to spend in its thinking process — a higher budget improves quality on harder tasks but increases latency and cost. Understanding when this trade-off is worth making is the key practical skill.

Task types where extended thinking reliably improves output quality

Multi-step mathematical problems: Any calculation requiring more than two or three sequential operations benefits from extended thinking. Claude catches arithmetic errors during its reasoning trace in a way that it cannot when producing answers directly.
Complex code debugging: When a bug involves multiple interacting systems, extended thinking lets Claude trace through the execution path systematically rather than pattern-matching to the most likely culprit.
Strategic planning and trade-off analysis: Tasks that require considering multiple competing options and their downstream consequences — architecture decisions, risk assessments, business analysis — benefit from the extra reasoning space.
Adversarial prompting resistance: On tasks where you need Claude to resist persuasion or identify flawed arguments (e.g. evaluating a business proposal with motivated reasoning baked in), extended thinking gives Claude room to notice the manipulation before responding.

When NOT to use extended thinking

Retrieval, format conversion, summarisation, and simple question-answering do not benefit meaningfully from extended thinking — and it adds latency and cost. Reserve it for tasks where the reasoning path itself is the hard part.

✅ Prompting for Reasoning Quality — Getting the Best Out of Claude's Think Step

Enabling extended thinking is not enough on its own — how you frame the problem significantly affects whether the reasoning process reaches a good conclusion. Claude's thinking scratchpad is exploratory and honest; it will follow whatever path the prompt opens up. A vague or underspecified prompt produces vague exploratory reasoning. A well-structured prompt that clearly defines what a correct answer looks like tends to produce focused, productive reasoning. Here are the structural choices that consistently improve outcomes.

Prompt structure for better reasoning

State the evaluation criteria upfront: Tell Claude what makes an answer correct or good before it starts reasoning. "A valid solution must (a) handle empty input, (b) run in O(n log n) time, (c) be readable by a junior engineer" gives Claude clear targets to reason towards.
Provide worked examples for complex domains: One or two examples of the reasoning style you want — showing the intermediate steps, not just the answer — dramatically improve the quality of Claude's own reasoning trace.
Ask for explicit uncertainty flagging: Instruct Claude to note in its final answer wherever its reasoning reached a step that felt uncertain. "If you are not confident about any step, note it in your response" produces more calibrated outputs than allowing confident-sounding answers on uncertain ground.
Set an appropriate token budget: For most tasks, 8,000–16,000 thinking tokens is a good starting range. Go higher (32,000–64,000) only for genuinely hard multi-step problems. Diminishing returns set in quickly beyond that for most practical tasks.

Claude's Daily Diary

Extended Thinking in Claude: When and How to Use Deeper Reasoning

✅ Extended Thinking — When Deeper Reasoning Pays Off

Task types where extended thinking reliably improves output quality

✅ Prompting for Reasoning Quality — Getting the Best Out of Claude's Think Step

Prompt structure for better reasoning