With extended thinking enabled, does adding "think step by step" to the prompt still help?

## Two ways to get a model to reason There are two distinct mechanisms that produce intermediate reasoning, and confusing them is a common interview stumble. **Prompt-level elicitation** works on any model: you ask for reasoning in the instructions, and the model writes it into the ordinary output before the answer. The reasoning lands in the visible reply unless you also impose an output contract that separates it. **Extended thinking** is a mode of the model and the serving stack: reasoning is generated on a separate channel before the reply, and post-training has specifically shaped how the model uses that channel. You do not ask for it in the prompt; you enable it in the request. Once the second mechanism is on, the first is largely redundant. The model is already doing the thing the instruction asks for, on a channel built for it. ## What redundancy actually costs It is not free to leave the boilerplate in. - **Tokens both ways.** The instruction occupies prompt space, and inviting reasoning into the answer channel inflates output length on top of the thinking you already bought. - **Contract leakage.** This is the failure that bites in production. A service expecting a JSON object gets a paragraph of deliberation followed by the object, because the prompt asked for step-by-step working and the model complied — in the visible channel. Parsers break; strict-output modes fight the instruction. - **Style collision.** Reasoning models are tuned to a particular way of using their scratchpad. Heavy prescription of a different reasoning style in the prompt can work against that tuning rather than with it, and the effect is task-dependent enough that only your eval set can tell you. ## What still helps Drop the ritual, keep the substance. Instructions that continue to earn their place on a reasoning model are the ones carrying information the model cannot derive: - **Domain constraints** — which regulation applies, which fields are authoritative, what a valid value range is. - **Task-specific checks** — "verify the totals reconcile before answering", "confirm every cited clause exists in the provided text". This is not "think step by step"; it names *what* to think about. - **Output shape** — the schema, the field names, the requirement that the reply contain no deliberation. On a reasoning model this instruction becomes *more* important, not less, because it is the thing keeping the answer channel clean. - **Worked examples** where the value is demonstrating a format or an unusual convention, not demonstrating that reasoning exists. A useful test: would this sentence tell a competent human colleague something they did not know? If yes, keep it. If it just says "be thorough", cut it. ## When you still reach for prompt-level reasoning Three cases keep it alive: 1. **Non-reasoning models.** Cheaper or older models have no thinking channel; instructions are the whole toolkit. Many production systems route the easy majority to such a model, and there the prompt does the work. 2. **You need the reasoning visible and structured.** If a downstream reviewer must see the working in a specific shape, and the provider only returns a summary or an opaque blob, generating it in the answer channel under an explicit format is the only way to get it. 3. **Thinking is disabled for latency.** On a latency-bound surface you may turn extended thinking off and accept a shorter, prompt-elicited reasoning pass because you can bound and stream it. ## The honest state of it This is empirically contested ground. Reports differ on whether explicit reasoning instructions on reasoning models are neutral, mildly harmful, or occasionally helpful on specific task families, and the answer has moved with each model generation. What has held up is the direction: as reasoning modes matured, generic CoT boilerplate went from essential to roughly inert, while precise task instructions stayed valuable. In an interview, say that, then say you would A/B it on your own eval set — one arm with the boilerplate, one without, same budget — and read accuracy, output-token count, and format-violation rate. That answer survives the next model release; a confident universal claim will not.