• AI Disruptor
  • Posts
  • The <scratchpad> technique will improve your AI outputs

The <scratchpad> technique will improve your AI outputs

Make your AI write out its "thoughts."

Welcome to AI Disruptor! if you want to join our growing community of readers, click the button below.

Today, we're diving into yet another powerful prompt engineering strategy that's going to take your AI interactions to the next level. It's all about using <scratchpad> tags to prime the AI with a roadmap for tackling your tasks.

When you're working with an LLM like ChatGPT or Claude, it's not just about what you ask - it's also about how you ask it. By taking a little extra time to guide the AI's thought process, you can unlock more structured, insightful, and targeted responses. That's where the <scratchpad> technique comes in.

While this technique should work with any LLM, I highly recommend using it with Claude. I’ve had the most success with it, and Claude is just all around better with writing tasks.

In this edition of AI Disruptor, you are about to learn:

Here's how the <scratchpad> technique works

  1. Start by giving the AI a clear task or objective, along with any necessary context or guidelines.

  2. But before you let the AI loose, add a section wrapped in <scratchpad> tags where you instruct it to generate a plan for approaching the task.

  3. In this scratchpad section, prompt the AI to break down the key steps, themes, or examples it will cover in its response. Essentially, you're asking it to show its work and walk you through its thought process.

  4. Review the AI's proposed plan and iterate as needed until you're confident it's on the right track.

  5. Once you give the green light, the AI will use that scratchpad outline to generate a beautifully structured and targeted response.

Example using the <scratchpad>

Let's see this in action with a real example. Let’s pretend AI Disruptor is a custom AI solutions provider and we want to write a blog about integrating company data with LLMs.

Here's what that prompt might look like:

I need to write a blog post on the topic of [integrating enterprise data with an LLM] for my AI solutions company, AI Disruptor.

Begin in <scratchpad> tags and write out and brainstorm in a couple paragraphs your plan for how you will create an informative and engaging blog. Also brainstorm how you will create a CTA at the end for our company, AI Disruptor.

And here is what your output from Claude will look like:

From here, you can either iterate on the scratchpad or move on to the outline, etc. The scratchpad planning process becomes even more powerful when you combine it with clear, detailed prompts that set the AI up for success.

Treat AI like it’s your colleague or collaborator

One of the coolest things about using <scratchpad> tags in your prompts is how it transforms your relationship with the AI. Instead of just hoping the AI magically reads your mind, you're working together to ensure the output aligns with your goals and expectations.

Think of the scratchpad as a shared whiteboard where you and the AI can brainstorm and refine your approach. By having the AI walk you through its thought process, you get to peek behind the curtain and catch any potential misunderstandings or missed opportunities early on.

This iterative back-and-forth is a really useful tool for prompt engineering. It allows you to nudge the AI in the right direction, provide clarifying guidance, and ultimately arrive at a response that's focused on your needs.

But to make the most of this collaborative power, you need to know how to effectively review and iterate on the AI's proposed plans. Here are a few key things to look out for:

🎯 Alignment with your main objective: Does the plan address the core goal you laid out in your prompt?

🔍 Coverage of key themes or points: Has the AI identified the most important aspects of the task to focus on?

🚦 Logical flow and structure: Does the outline present a clear, coherent narrative arc or problem-solving approach?

💡 Creativity and insights: Has the AI brought any unexpected ideas or angles to the table that could enrich the response?

When reviewing the scratchpad, don't be afraid to get specific with your feedback. Call out what you like, what you think could be improved, and any additional points you want the AI to consider. The more targeted your guidance, the better the AI can adapt and refine its approach.

Some people like to make prompt engineering or communicating with an LLM look like a super-complex process. While sometimes this is the case, other times it’s very simple. Just talk to the AI like you are talking to a colleague. Don’t overthink it.

And if the first attempt doesn't quite hit the mark, don’t worry. The beauty of the <scratchpad> technique is that you can keep iterating until you're fully satisfied. It's a collaborative process of evolution and elevation.

Once the scratchpad plan gets your stamp of approval, sit back and let the AI work its magic. Thanks to all the groundwork you laid together, you can trust that the full response will be a lot more aligned.

There you have it. By embracing the <scratchpad> technique, you'll supercharge your AI interactions, generate more targeted and insightful content, and achieve results that help set you apart.

Start experimenting with <scratchpad> tags in your own prompts and witness the magic firsthand. Once you experience the power of a structured, iterative approach, you'll never go back to the old way of prompting.

- Alex (Creator of AI Disruptor)

What did you think of this edition of AI Disruptor?

Your feedback helps me create a better newsletter for you!

Login or Subscribe to participate in polls.


or to participate.