Effective prompting techniques

Introduction to Effective Prompting

Let's explore one of the most practical skills when working with AI: crafting effective prompts. This might sound technical or complicated, and some guides certainly make it seem that way, but at its heart, it's surprisingly straightforward. Prompting is simply how we apply this course's "Description Competency" in practice—clearly communicating what we want, how we want it done, and how we want to interact with our AI assistant throughout the entire process.

Think of prompting like explaining a task to a helpful new colleague who's eager to assist but needs clear directions and expectation setting to do their best work. We'll be using Claude throughout this section, but these tips can be carried over to many other AI systems.

What is Prompt Engineering?

You might have heard the term "Prompt Engineering" tossed around. Prompt Engineering is simply the practice of designing effective instructions for AI systems like Claude. It's about crafting your questions and providing context in ways that help AI assistants understand exactly what you want.

What's fascinating is that effective prompting blends familiar human communication skills with a few considerations specific to AI. Many principles that make for good human conversation—such as being clear, providing relevant context, and giving concrete examples—also apply when working with AI. Yet there are differences, such as being more explicit about things humans could naturally infer, accommodating the AI's limited context window, and sometimes, depending on the AI you're working with, using specific formatting that machines can easily process.

As AI assistants continue to evolve, prompting best practices evolve too. What works with today's AI systems may be different from what works with tomorrow's. Experimentation is key to discovering what works best for your specific needs.

Six Foundational Prompting Principles

In this video, we'll mainly explore six foundational prompting tips that will go a long way toward helping you effectively communicate and collaborate with Claude and other AI systems:

1. Give Claude context.
2. Show examples of what "good" looks like.
3. Specify output constraints.
4. Break complex tasks into steps.
5. Ask Claude to think first.
6. Define Claude's role, style, or tone.

1. Provide Context

The first principle is simple but powerful: be specific and clear about what you want, why you want it, and perhaps most surprisingly, who you are.

Let's take a simple prompt: "Tell me about climate change." How can we improve this by giving Claude more context? A more specific, context-rich version could look like: "Explain three major impacts of climate change on agriculture in tropical regions, with examples from the past decade."

Our baseline prompt was vague and left Claude guessing about our interests, level of knowledge, and the depth of detail we're looking for, such as geography and time span. We can even add more context by providing information not just about what we're looking for, but why we're asking and how we'll be using that information.

Now our prompt looks like this:

"Explain three major impacts of climate change on agriculture in tropical regions, with examples from the past decade. I'm preparing for a job interview at an agricultural research lab in Indonesia. I have a degree in ecology but no specific knowledge on climate change. Write a summary of key concepts that would help me speak intelligently in the interview."

All this added context helps tailor Claude's response to your specific situation and knowledge level. This kind of background information is something we naturally provide in human conversations but might forget to include when talking with Claude.

2. Show Examples (Few-Shot Prompting)

Sometimes showing is better than telling. Providing examples of the kind of output you're looking for can be incredibly effective. This is sometimes called "Few-shot prompting" or "N-shot prompting" in technical circles, where n is the number of examples given.

For instance, take the following prompt: "Please convert this technical statement to plain language: 'The platform implements end-to-end encryption protocols to safeguard data integrity.'" Claude may already be able to do this to your satisfaction, so we recommend you just try first without examples.

But let's say you have a very specific style you want Claude to follow, and it's harder to explain than to give examples. Your refined prompt could look like this:

"Here are two examples of how to convert technical jargon into plain language: Original: The quantum algorithm exhibits quadratic speedup. Plain: The new method solves problems roughly twice as fast as previous methods. Original: The interface leverages intuitive design paradigms. Plain: The design is easy to understand and use. Now, please convert this complex technical manual to plain language."

When providing examples, aim to cover the full diversity of possible prompts, such as examples that cover different cases or styles. This helps Claude better understand the broad range of the pattern you want it to follow.

3. Specify Output Constraints

Being clear about output constraints—such as the desired format and length, the language you want Claude to code in, or the color of the buttons on a webpage—ensures you get exactly what you need.

Example:

"Create a clean, modern, single-page art portfolio website. Include these main sections: Hero, About Me, Skills, Portfolio Projects, Experience, and Contact. Make the navigation menu sticky and responsive with a hamburger menu on mobile. Use a sunset color palette and add a dark/light mode toggle in the navigation."

Guidance like this helps Claude structure its response to match your expectations.

4. Break Complex Tasks into Steps

When you have a complicated request, breaking it down into smaller steps helps Claude follow your thinking and deliver better results. This is sometimes called "Chain of Thought" prompting.

Instead of asking Claude to "Analyze this quarterly sales data," you might say:

"I'd like to analyze this quarterly sales data. Please approach this by:

1. Looking through our sales records to identify the top-performing products.
2. Comparing current quarter results to the previous quarter.
3. Highlighting any unusual trends or patterns.
4. Suggesting possible reasons for these trends."

Listing out task steps ensures that Claude follows the process you want. While modern reasoning models are increasingly capable of performing step-by-step reasoning on their own, you should still guide the process when the task relies on your specific domain expertise.

5. Ask Claude to Think First

Sometimes it can be helpful to explicitly give AI assistants like Claude space to work through its process first before executing its task. This approach helps Claude produce more thorough and well-considered responses.

For example, you can add this to your prompt: "Before answering, please think through this problem carefully. Consider the different factors involved, potential constraints, and various approaches before recommending the best solution."

Modern reasoning or "Extended Thinking" models by default think before acting. However, if you're working with a model that does not, you can still prompt it to do so. It is important to give the AI space to think before doing its task, not after. This also allows you to see where the AI might be going astray, helping you hone your Description Competency further.

6. Define Claude's Role, Style, or Tone

Specifying how you want Claude to communicate can significantly change how it approaches a task. Simply put: who do you want the AI to act as?

Example: "Please explain how rainbows form from the perspective of an experienced science teacher speaking to a bright 10-year-old who's interested in science."

This is also a good way to brainstorm or get feedback. You can specify a general role or even ask Claude to take on the persona of a specific figure, such as Richard Feynman. Example: "As a UX design expert, review this website wireframe and suggest three improvements focusing on user navigation and accessibility."

Iteration and Refinement

Perhaps the most powerful technique is asking Claude to help improve your prompt. When you're not sure how to ask for something, describe your situation and ask: "I'm trying to get you, Claude, to help me with [Goal]. I'm not sure how to phrase my request to get the best results. Can you help me craft an effective prompt for this?"

Effective prompting is iterative and experimental. Your first attempt won't always yield the perfect result. When a response isn't quite what you need, try:

• Adding more specificity or context.
• Providing examples of your desired output.
• Breaking the task into smaller steps.
• Asking for variations: "Can you give me three different versions of this?"
• Requesting different formats: "Instead of a paragraph, could you present this in an interactive Artifact?" (Note: Artifacts are a unique way Claude creates easy-to-digest outputs).
• Checking confidence: "How confident are you about this answer?"
• Resetting the conversation: Sometimes starting fresh gives better results than trying to correct a conversation that has gone off track.

Summary of Best Practices

As you apply these techniques, here is a recap of patterns that consistently work well:

• Start with a clear task overview statement.
• Include format specifications and examples.
• Set explicit constraints or requirements.
• Provide rich and relevant background information.

Common mistakes to avoid:

• Assuming Claude can read your mind.
• Overloading a single prompt or conversation with multiple unrelated tasks.
• Being too vague about what success looks like.
• Not providing feedback on previous responses.

Effective communication with AI systems like Claude combines timeless human communication principles with AI-specific techniques. These six principles, together with the "secret weapon" of asking Claude for help, form a solid toolkit for your AI interactions. Maintain a spirit of experimentation and adapt your approach based on your results.