How to Write Better DeepSeek Prompts for Reasoning and Search

The short version

Say everything up front, pick the right switch, and say what the answer should look like. DeepSeek does its best work when it has the full picture before it starts reasoning.

Step by step

1. Pick the mode. Quick tasks: no switches. Problems with steps: DeepThink. Anything recent: Search. See use DeepThink.
2. Give context. Who you are, who the result is for, what you already know.
3. State the goal in one sentence.
4. List constraints: budget, length, deadline, rules that must hold.
5. Set the format: table, bullet list, numbered steps, word limit.
6. Ask it to check. “Before you answer, check the result against every constraint.”

What’s different with DeepSeek

• Don’t over-script reasoning. With DeepThink on, DeepSeek works out the steps itself. Describe the problem precisely instead of dictating the method, unless you need a specific method (for school, say).
• One complete message beats many small ones. Reasoning starts from what’s in front of it.
• Local R1 models: DeepSeek’s model card recommends no system prompt, all instructions in your message, and a temperature of 0.5 to 0.7.
• For structured output in the API: DeepSeek’s JSON Output mode needs the word “json” in the prompt and an example of the format you want.

Prompts to copy

The all-purpose template:

Context: [who you are, the situation]
Goal: [one sentence]
Constraints: [list]
Format: [table / bullets / under X words]
Before answering, check your answer against each constraint and fix anything that fails.

A plan (DeepThink on):

I'm opening a coffee cart at weekend markets. Budget $8,000, launch in 10 weeks,
I work weekdays. Make a week-by-week plan with costs. Flag the riskiest step and why.

Research (Search on):

Find the current requirements for [permit / rule] in [city, state].
Give a short answer, then a list of official sources with links and dates. Say if anything is unclear.

Get it to ask first:

I want help writing [thing]. Before you start, ask me up to five questions
you need answered to do it well.

Example

Weak prompt: “Help me with a budget.”

Strong prompt (DeepThink on): “Monthly take-home $4,200. Rent $1,450, car $380, groceries about $600. I want to save $500 a month and pay off a $3,000 card at 22% within a year. Build a monthly budget table and tell me if it’s realistic. If not, what’s the smallest change that makes it work?”

Why it works: every number is there, the goal is clear, and it asks for a verdict, not just a table.

Mistakes to avoid

• Vague goals. “Make it better” gives DeepSeek nothing to aim for.
• Turning on every switch. DeepThink and Search add time; use them when they help.
• Long chats on unrelated topics. Start a new chat when the subject changes.
• Real names and numbers you don’t need. Use placeholders. See keep your data private in DeepSeek.

Good to know

If a reply goes off track, a clearer version of your question usually works better than piling on corrections. DeepSeek’s chats can have branches: its FAQ says that if part of a chat seems to have disappeared, the arrow buttons (with numbers) below a question or answer switch between branches.

Related: use DeepThink and use DeepSeek Search.