Reasoning Model Discipline

به چه درد می‌خورد

می‌گوید هنگام بردن یک پرامپت به مدل استدلالی چه چیزهایی را باید **حذف** کنی: شخصیت‌سازی، دستور «گام‌به‌گام فکر کن»، و مثال‌های few-shot. و یک چیز اضافه کن که مدل‌های گفتگویی نادیده می‌گیرند: حالت شکست مورد انتظار.

خود اسکیل کپی

A prompt tuned for a chat model is usually a bad prompt for a reasoning model.
Strip it down.

Remove:
- The persona ("You are a world-class expert..."). It competes with the model's
  own reasoning and buys nothing measurable.
- Chain-of-thought instructions ("think step by step", "reason carefully").
  The model already does this; asking again makes it narrate rather than think.
- Few-shot examples, unless the FORMAT is unusual. Examples anchor a reasoning
  model to the shape of the example instead of the shape of the problem.

Keep, and make precise:
- The task, in one unambiguous sentence.
- The constraints, as a list, each one checkable.
- The output format, exactly.
- The definition of done — how you will decide the answer is correct.

Add one thing chat prompts rarely need:
- **The failure mode you expect.** "The common wrong answer here is X, because
  Y." A reasoning model uses that; a chat model mostly ignores it.

Keep the temperature low and the output budget high. Truncating a reasoning
model mid-derivation produces confident nonsense, not a shorter answer.

نصبش چطور است کپی

DeepSeek

1. Start a new chat and paste the text below as the first message.
2. For the API, send it as a `system` role message.

With the reasoning model, keep the instructions short and factual:
long persona preambles compete with the model's own reasoning.

برای تو جواب داد؟

از کجا آمده