Home / Writing a request that works
Making Claude reason before concluding
Asking Claude to write out its reasoning before its conclusion improves accuracy on multi-step questions, whether that reasoning is invited by a sentence in the request or is already active as a standalone mode, and both versions consume extra tokens.
On a multi-step question, a budget to split across three spending categories for instance, an immediate response often skips a constraint. Asking Claude to write out the steps of its reasoning before the conclusion reduces this risk: each step is exposed separately, which makes it possible to reread it and spot an error before it carries through to the final response.
Two ways to get this reasoning
The first is to write it into the request itself, a sentence such as think step by step before giving your final answer is enough:
Une association reçoit 850 euros pour trois postes de dépense,
fournitures, transport, communication. Le poste transport doit
rester sous 200 euros. Réfléchis étape par étape avant de donner
ta répartition finale.Anthropic's documentation presents this wording as a fallback for when automatic reasoning is not active. The second way relies on a reflection mode built into Claude, switchable in the settings, where the model itself decides how much reasoning depth is needed. Checked on 2 September 2026: on the Claude Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5 models, Anthropic's technical documentation states that this reasoning stays active permanently, in adaptive mode. The help centre aimed at people using the app specifies, for Claude Fable 5.1 and Claude Opus 5, that this mode cannot be switched off.
What this costs
Each extra reasoning step adds to the text Claude produces, and that text is counted in tokens, the fragments Claude processes and is billed for. Anthropic's help centre puts it this way: a higher reflection level produces more thorough responses, but they take more time and consume more tokens, which brings usage limits closer sooner. Reasoning invited by a simple sentence in the request costs the same way, since the written steps also add to the text produced.
Matching the length of the reasoning to the task
A simple question, a date to calculate, a word to define, does not need written steps: the token cost then buys nothing. A question that combines several constraints, as in composing a complex request without inventing, benefits instead from explicit reasoning, because each constraint can then be checked one by one before the conclusion.
Reasoning written out before the final response
A volunteer coordinator asks Claude to split forty seven chairs across three rooms, whose maximum capacities differ, adding to the request an instruction to write out the reasoning step by step before giving the final split. Claude answers with five numbered steps, followed by a final allocation table.
Write, in one sentence, what this situation establishes, and in one sentence what it does not establish.
What this establishes: This response establishes that the step-by-step reasoning instruction produced a sequence of five numbered steps, visible before the final table, on this specific request.
What this does not establish: It does not establish that the final chair allocation actually respects the maximum capacities of the three rooms, which only checking the table would confirm.
The three most common miscalibrations
- Too broad This instruction now guarantees that any Claude response on an allocation problem will be accurate, whatever the number of constraints involved.
- Too narrow This result proves nothing, since the request only concerned a single case of allocating chairs across three rooms.
- Off target This result shows that the maximum capacities of the three rooms had been correctly measured before the request.
- Writing the reasoning steps before the conclusion, in the text of the request, reduces the risk that a constraint gets forgotten on a multi-step question.
- On the Claude Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5 models, this reasoning stays active permanently and cannot be switched off.
- Each extra reasoning step, whether invited by a sentence or produced by an active mode, consumes extra tokens.
- A simple question does not need written steps, the token cost buys nothing as long as no constraint risks being forgotten.
Open a conversation with Claude on a question that needs several steps of calculation or logic, for example splitting a budget across three spending categories. Add to your request the sentence think step by step before giving your final answer, send the request, then reread the steps written before the conclusion.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Prompting best practices consultée le 2026-09-02
- Changing the model, effort and reasoning settings consultée le 2026-09-02