Concept L3 · Workflows informational

Let Reasoning Models Think Before They Answer

What it is

A mental model for prompting ChatGPT's reasoning models: give them a well-defined goal and the constraints, then get out of the way — they do their own step-by-step internally.

Fights the modelover-steering· 'Think step by step'· Rigid format demands· Interrupting mid-taskHelps the modelclear target→ State the goal + constraints→ Give success criteria→ Let it work, then check
Reasoning models reward a clear target and room to work.

Why it works

Reasoning models are trained to work through a problem before answering. Telling one to 'think step by step' or forcing a rigid shape can actually get in the way. What they need is an unambiguous target and clear success criteria.

When to use it

Hard, multi-step problems: proofs, planning, tricky debugging, trade-off analysis. Anywhere a single wrong assumption early would sink the whole answer.

When not to use it

Simple, fast tasks — a reasoning model is slower and no better at 'rewrite this politely'. Match the model to the job as a ChatGPT basics skill.

Prompt

Goal: <precise objective>.
Constraints: <hard limits>.
Done looks like: <how I'll judge the answer>.

Work the problem however you need, then give me the answer and a short note on the key assumption it rests on.

Example

Asked to design a fair on-call rotation for six engineers with three constraints, the reasoning model returns a schedule plus the one constraint that forced the trickiest trade-off.

Advanced version

When the stakes are high, ask for the answer plus the single input that, if wrong, would most change it. That points your verification at the weakest joint instead of re-checking everything.

Common mistakes

  • Micromanaging the model's process instead of specifying the outcome.
  • Omitting success criteria, so 'done' is whatever the model decides.
  • Using a reasoning model for tone or formatting, where it adds latency and nothing else.