Prompting DeepSeek V4 Pro and Flash: A Practical Guide

DeepSeek V4 is a text-only model family with a context window of roughly one million tokens, strong reasoning, and solid coding skills. Both versions are available in GPTunneL: the flagship DeepSeek V4 Pro and the fast, lightweight DeepSeek V4 Flash. Both can reason step by step, call tools and functions, and search the web — and the DeepSeek family remains one of the most affordable options in its class. This guide shows how to get the most out of them with well-built prompts.

Pro or Flash: how to choose

  • V4 Pro — for hard problems: multi-step logic, code architecture, deep analysis of long documents, anything where mistakes are costly.
  • V4 Flash — for fast, high-volume work: drafts, summaries, small code fixes, classification, routine questions.

A practical strategy: start with Flash. If the answer feels shallow or the logic falls apart, rerun the same prompt on Pro. You spend the extra compute only where it pays off.

What to keep in mind

  • The model is text-only. It cannot see images, screenshots, or diagrams. Describe anything visual in words: "a table with three columns: date, amount, status" instead of an attached screenshot.
  • Reasoning is its strength. Don't bury logic problems under "give me a short answer" — let the model think step by step and demand brevity only from the final conclusion.
  • Code is the other strength. Specify the language, version, style, and your definition of done: tests pass, errors are handled, docstrings exist.
  • Context of ~1,000,000 tokens. You can load an entire codebase or a stack of documents at once. Structure still matters: label your materials with headings and ask the model to reference specific sections.

Techniques with examples

Prompt structure

Role → task → context → output format. Works equally well on Pro and Flash:

You are a technical writer. Task: rewrite the instructions below for beginners. Context: the reader is seeing the product for the first time. Format: numbered steps, one action per step, no jargon.

Chain-of-thought for math and logic

Explicitly ask for the reasoning first and the answer last:

Solve this problem step by step. First list the known facts, then a solution plan, then the calculations with a check after each step. Put the final answer on its own line at the end.

Generating and reviewing code

Set boundaries for generation, criteria for review:

Write a Python 3.12 function that parses CSV with date validation. Requirements: type hints, malformed rows handled without crashing, a docstring with an example. After the code, add 3 pytest unit tests.

Review the code below. Check in this order: bugs and edge cases, security, performance, readability. For each finding, point to the exact line and suggest a fix. Do not rewrite the whole file.

Few-shot: show a sample

Two or three input → output examples lock in a format better than any description:

Give feedback on the claim, in the style of the examples. Claim: "Coffee is bad for you" → Feedback: "Too absolute: specify the dose and the group of people." Claim: "Everyone should learn Python" → Feedback: "Debatable: depends on the profession, add a condition." Claim: "Remote work is more productive" →

Working with long context in stages

A million tokens is not a reason to dump everything into one question. Break it up:

  1. Load the materials and ask for a map: "Build a table of contents for these documents with a one-line summary of each section."
  2. Work section by section: "Now, using section 3 only, list every obligation of the parties."
  3. Consolidate: "Merge the conclusions from the steps above into a single report and flag contradictions between sections."

Ready-made prompts — copy and adapt

You are an analyst. Read the documents below and answer this question: [question]. Quote the passages you rely on and name the section. If the documents don't contain the answer, say so — don't guess.

Refactor the function below while preserving its behavior. First list what you will improve and why, then give the new code, then what to verify before merging.

Explain [topic] on three levels: 1) one paragraph for a beginner, 2) a detailed walkthrough with examples, 3) subtleties and common mistakes. Use a heading for each level.

Use web search to find recent information on [topic]. Compile a briefing: what has changed, the source behind each claim, and what remains unconfirmed. Separate facts from speculation.

Draft a project plan: [description]. Format: phases with goals, tasks within each phase, risks with mitigation options, and done criteria per phase. Finish by asking me 3 clarifying questions.

Troubleshooting

  • The answer feels shallow — you're on Flash with a hard problem: switch to Pro or explicitly request step-by-step reasoning.
  • The model "sees" a nonexistent image — it's text-only; replace the attachment with a verbal description or the extracted text.
  • Details get lost in a long document — don't ask for everything at once: build a content map first, then question it section by section.
  • The code comes back in the wrong style — the boundaries weren't set: name the language, version, and linter, and paste a sample of your own code as a few-shot example.
  • The answer is confident but suspicious — add "if you're not sure, say so," ask it to re-check its reasoning, or enable web search to verify the facts.

Both models are available in GPTunneL on a single balance: run the same prompt on Flash and on Pro — the difference will tell you which tasks belong to which model.

Try it in GPTunneL