Unlocking the full potential of advanced AI systems like OpenAI’s o1 and DeepSeek R1 requires a new approach to prompting. Forget what you thought you knew about traditional prompt engineering—this guide reveals what works (and what doesn’t) with today’s reasoning models.

The Evolution of Prompting: Why Things Have Changed

The landscape of AI is rapidly evolving. Older large language models (LLMs) often relied on intricate prompt engineering techniques to guide their responses. However, newer “reasoning models” are fundamentally different. They’re designed to perform complex internal thinking, which drastically alters how you should structure your requests. Tyler Fyock’s research highlights this shift, revealing that many of the strategies that once boosted performance with earlier models now hinder the effectiveness of these advanced systems.

Part 1: Prompts That Unlock Reasoning Power

To leverage the capabilities of modern reasoning models effectively, it’s crucial to adopt a minimalist and direct prompting style. Let’s dive into seven prompt types that consistently deliver strong results. We’ll also provide examples to illustrate each approach.

1. Direct, Clear Questions & Instructions

The cornerstone of successful prompting with reasoning models is clarity. Avoid ambiguity and get straight to the point. Instead of hinting at what you want, explicitly state your request. This eliminates any potential for misinterpretation and allows the model to focus on fulfilling the task at hand.

• Example: Instead of “Tell me about this article,” use “Summarize this article in three sentences.”
• Example: Instead of “What are some arguments about X?”, use “List the main arguments for and against X.”

2. Specific Constraints and Boundaries

Reasoning models excel when given clear parameters. Setting boundaries – whether they relate to budget, audience, or format – focuses the model’s response and improves accuracy. Specifying the desired outcome acts as a crucial guide.

• Example: “Propose a solution with a budget under $500.”
• Example: “Explain this concept for a 10-year-old.”
• Example: “Write a short story under 500 words focusing on the theme of resilience.”

3. Zero-Shot Prompts: Letting the Model Shine

Modern reasoning models are frequently optimized for “zero-shot” performance, meaning they can tackle tasks without being shown example inputs and outputs. Avoid the temptation to provide multiple examples; instead, simply describe the task and desired output directly.

• Example: “Translate the following English text into French: ‘The quick brown fox jumps over the lazy dog.’” (No examples of English-French translations are given).

4. Delimited Inputs for Enhanced Clarity

When including multiple pieces of information or instructions within a single prompt, using delimiters is incredibly helpful. This visually separates the different components and prevents confusion. Common delimiters include markdown, XML tags, or simple section titles. This helps the model parse the information correctly.

• Example: “Here’s the article: [Article Text Here]. Summarize the key arguments in five sentences.”
• Example: “Instructions: Write a marketing email. Subject: Special Offer. Body: [Email Body Text Here]”

5. Task-Oriented Prompts: Focus on the Goal

Instead of focusing on the process, direct the model towards the desired outcome. Frame your requests as specific tasks that need to be completed. This helps the model understand the ultimate purpose of its response.

• Example: “Draft an email to request a meeting.”
• Example: “Solve this math problem: 2x + 2 = 10.”
• Example: “Create a list of five blog post titles about sustainable living.”

6. Iterative Refinement: Guiding the Model to Perfection

Sometimes, the initial response isn’t quite right. Instead of providing more examples or lengthy explanations, encourage the model to keep reasoning and iterating until it meets your criteria. This leverages its internal reasoning abilities to refine its output.

• Example: “Keep revising this summary until it’s under 100 words and covers all key points.”
• Example: “Refine this draft until it is grammatically perfect and accurately reflects the information.”

7. High-Level Guidance: Steering, Not Micromanaging

Avoid micro-managing each step of the process. Provide a big-picture goal and allow the model to determine the optimal approach. This fosters creativity and leverages its ability to find innovative solutions.

• Example: “Develop a plan to increase productivity in a remote team.” (Instead of specifying each task).
• Example: “Create a strategy to expand our company’s presence on social media.”

Part 2: Prompts to Avoid – and Why They Fall Short

Now that we’re clear on what works, let’s examine prompting techniques that are ineffective, or even detrimental, when working with modern reasoning models. These methods were often useful in the past but are no longer optimal.

1. Chain-of-Thought (“Think Step by Step”) – A Relic of the Past

Older LLMs often struggled with complex reasoning and benefited from prompts that encouraged them to “think step by step” or “explain your reasoning.” Modern reasoning models *already* perform this internal reasoning. Asking them to explicitly detail their thought process adds unnecessary complexity and can actually degrade performance.

2. Few-Shot Prompts – The Overwhelming Effect

Providing multiple examples of input/output pairs, known as “few-shot prompting,” often overwhelms reasoning models. They’re designed to perform well with just a task description, and providing examples can confuse them and lead to less accurate results.

3. Overly Complex or Long Prompts – The Information Overload

Brevity is key. Too much context or unnecessary details can overwhelm the model and make it difficult to discern the core request. Keep prompts concise and focused on the essential information.

4. Ambiguous Instructions – The Path to Generic Responses

Vague or open-ended prompts often lead to generic or off-target responses. Ensure your instructions are clear, specific, and have a defined objective.

5. Structured Output Demands (Unless Necessary) – Limiting Reasoning

Forcing the model into rigid templates or formats can stifle its reasoning ability and result in less useful answers. Allow it the flexibility to generate the most appropriate response, unless a specific format is absolutely required.

6. Excessive Role-Playing – Distraction from the Task

While some role-playing prompts can be useful, overcomplicating the request with multiple personas or scenarios can distract from the core task and reduce performance.

7. Redundant Prompts – Clarity Through Simplicity

Repeating the same instructions or information within a prompt doesn’t help and can actually confuse the model. Strive for clarity and precision in your initial request.