OverpayingForAIPricing desk

Lesson 5 of 8 · 10 min read · Beginner → Intermediate

Prompting basics for DeepSeek

How to write prompts that work with DeepSeek's chat and reasoner models, with three before-and-after examples and the token habits that keep API costs down.

In this lesson you will

  • Structure a prompt so DeepSeek does the right task on the first try
  • Adapt prompts for the reasoner model, which wants less hand-holding
  • Rewrite three weak prompts into strong ones
  • Avoid prompt habits that quietly multiply token costs

DeepSeek's chat models respond well to the same structure that works everywhere: say who the model is acting as, give it the material, state the task precisely, and describe the output you want. The reasoner is the one place where the advice differs — it does best with a clear problem statement and minimal step-by-step instruction, because it will generate its own steps.

Anatomy of a prompt that works first timeFive parts. Most bad answers are missing two of them.RoleWho the model should be
“You are a pricing analyst…”
TaskThe one thing you want
“Rank these 4 plans by cost per 1K tokens.”
ContextFacts it cannot guess
paste the plan table, your usage
FormatShape of the answer
“a table, then a 2-line verdict”
ConstraintsWhat to avoid
“no marketing language; cite the source”
Cost angleA precise prompt costs a few hundred tokens more and saves the two or three retries that cost thousands.Asking for a short format (“3 bullets”) cuts output tokens — the expensive kind — by half or more.
Figure 1.The four parts of a working prompt. With DeepSeek, the context section is where token cost lives; the task and format sections are where quality lives.

The four parts

  • Role. One line. "You are a technical editor for a SaaS company." It sets tone and assumptions cheaply.
  • Context. The material: the draft, the data, the code, the constraints. This is the bulk of your input tokens, so include what is needed and nothing more.
  • Task. A verb and an object. "Rewrite", "list", "find the bug in", "compare". Vague tasks produce vague answers and long ones.
  • Format. What the answer should look like: length, structure, language, whether to include reasoning. This is the single biggest lever on output tokens.

Example 1: writing

Before: "Make this email better." The model has no idea what better means, so it guesses, usually by making it longer and more formal.

After: "You are a plain-English editor. Below is an email to a customer whose order is delayed. Rewrite it to under 120 words, keep the apology and the new delivery date, remove all jargon, and end with one clear next step for the customer. Return only the email." Same input, but now the model knows the audience, the length, what to keep and what to return.

Example 2: coding

Before: "Why doesn't this work?" followed by 400 lines of code. The model will summarise the whole file and speculate, which costs output tokens and rarely finds the bug.

After: "This Python function should return the median of a list but returns the wrong value for even-length lists. Here is the function and a failing test. Find the bug, explain it in two sentences, and return the corrected function only." Paste the 20-line function, not the file. If the chat model fails, switch to the reasoner with the same prompt — this is the escalation pattern from Lesson 2.

Example 3: analysis

Before: "Analyse this data." A pasted CSV and a one-word instruction produces a generic tour of the columns.

After: "Here is monthly revenue by region for 24 months as CSV. Identify the three regions with the largest change between the first and last six months, give the percentage change for each, and flag any month that looks like a data error. Answer as a table plus at most three bullet points." Now the output is short, checkable, and useful.

Prompting the reasoner differently

For deepseek-reasoner or DeepThink mode, drop the "think step by step" instructions — the model does that itself, and telling it to can produce redundant reasoning and more tokens. State the problem precisely, give all constraints up front, and ask for the final answer in a specific format. Avoid long system prompts; the reasoner responds best to a short, clean user message.

Knowledge check

You are sending a maths word problem to deepseek-reasoner. Which prompt addition is most likely to waste tokens?

Lesson FAQ

Does DeepSeek support system prompts?

Yes. The API accepts a system message in the standard OpenAI-compatible format. The chat app does not expose a custom-instructions field the way ChatGPT does, so put persistent instructions at the top of each chat.

Should I prompt DeepSeek in Chinese or English?

Both work well. Use whichever language your material and audience use; the model handles both natively and translation-in-the-middle only adds tokens.

Finished reading?

Mark it done to track your progress through the course.

Compare, calculate, decide — for DeepSeek

If our calculators helped you cut down on hidden AI wallet leaks, consider buying us a coffee. A tiny fraction of your savings keeps our pricing indexes updated daily.

Not sure which AI is cheapest for your use case? Find out in 30 seconds — no signup required.

AI cost intelligence

Stop overpaying for AI tools

Join the OverpayingForAI list for pricing updates, cheaper alternatives, and practical buying guidance.

Now tracking 50+ AI tools, models, platforms, subscriptions, coding tools, and automation products.

We use your email only for OverpayingForAI updates. Unsubscribe anytime.