Exponential Productivity: AI to the Nth Power

This Action-Powered Newsletter will be slightly different—and delightfully concise! Instead of sharing concepts or giving advice, I’m going to give you a glimpse of something that I’m working on right now and explain what it’s taught me about collaborating with AI.

Large Language Models (LLMs) like ChatGPT or Claude are valuable even when all you’re doing is drafting emails or modifying recipes. For example, I’ve gotten more work done on our family cars the past three months than I ever did before, all because quick conversations with ChatGPT guided me.

But when you use AI to develop other AI tools, something magical happens. There’s a virtuous feedback loop that compounds your efforts. I’ve seen this firsthand while developing “Axton,” the AI Productivity Assistant for attendees of my upcoming Fundamentals of Action-Powered Productivity course.

The Compounding Effect of AI Collaboration

A little background: Axton is built on OpenAI’s Assistants API, and is specifically designed to help my clients implement the Action-Powered Productivity (APP) approach. Unlike general AI assistants that know a little about everything, Axton is deeply focused on helping people implement the productivity principles I’ve developed over the past decade.

I’m using AI in three primary ways as I build Axton.

1. Claude helps me code

Axton’s chat interface is built with PHP, HTML, and Javascript, and it’s styled with CSS. I’ve coded using all of those before, but my skills are rusty — and, honestly, I’ve never been more than an intermediate PHP coder or a novice Javascript coder.

By collaborating with Claude, I dramatically increase my coding skill. Notably, the Axton app is using the Alpine.js framework — which I do not know anything about. It’s grounded in Javascript, so I can parse it enough to get by, but I would struggle if I were coding with it from scratch.

Axton's chat interface combines clean design with contextual responses. The interface you see was built with Claude's help, while the responses come from GPT-4o-mini guided by carefully refined prompts. — Axton’s chat interface combines clean design with contextual responses. The interface you see was built with Claude’s help, while the responses come from GPT-4o-mini guided by carefully refined prompts.

Before you declare that AI is coming for everyone’s livelihoods, let me point out: Claude is a powerful collaborator, but I also bring elements to the table. I can read the programming languages we’re using. I know how to load files onto my web host’s servers. I know how to access and use databases on my web host — though Claude’s SQL code was a major time-saver. Claude accelerates my work by orders of magnitude, but it isn’t collaborating in a vacuum. I’m a competent partner.

2. The OpenAI Assistants API provides Axton’s “brain”

Axton on mobile, ready to help course participants implement Action-Powered Productivity principles. The responsive design ensures a seamless experience across all devices.

The OpenAI Assistants dashboard allows me to provide documents for Axton to access when responding to user inputs. I uploaded two versions of my book The Rhythms of Productivity: an abridged copy for handling simple questions, and a full version for more detailed inquiries.

Axton has access to a document that breaks down a variety of productivity apps and how they can fit into an Action-Powered Productivity approach, as well as a collection of my other writings. Plus, there are toggles for the “creativity” of AI responses that I can calibrate through experimentation. I enjoyed the utter chaos when I had the “Temperature” and “Top P” at their highest settings, but I pulled them back a bit.

And none of this even mentions the actual prompt, which is where the “compound interest” of pointing AI at AI really pays dividends.

3. ChatGPT helps me refine Axton’s prompt

Behind the scenes: OpenAI's Assistants dashboard, where I fine-tune Axton's behavior. The prompt you see here is just the beginning—it continues for another three or four screens’ worth, thanks to iterative refinement with ChatGPT. — Behind the scenes: OpenAI’s Assistants dashboard, where I fine-tune Axton’s behavior. The prompt you see here is just the beginning — it continues for another three or four screens’ worth, thanks to iterative refinement with ChatGPT.

Refining Axton’s prompt can be like playing Whac-A-Mole, and that is only going to intensify when it isn’t just me doing the testing. But there’s an amazing shortcut available: use AI to improve AI.

When I get a response from Axton that I’m not fond of, I head to ChatGPT. I share the current prompt and the interaction I didn’t like, and dictate my concerns. Then I ask it to assume the role of the AI executing the prompt — that is, Axton — and provide feedback on the prompt from that perspective. After that, I ask it to rewrite the prompt based on its own feedback. This virtuous feedback loop is like “AI squared.”

There’s one more layer still. I can ask ChatGPT to generate test questions designed to poke and prod at Axton’s weaknesses. I could do this on my own, but it’s a lot faster and easier to let AI try to “break” itself. Now we’re up to “AI cubed,” and the results I’ve seen are more than encouraging — they’re exhilarating!

I’m sure creating and refining prompts for AI Assistants will get even more efficient as I get comfortable using OpenAI’s built-in back-end tools. But conversing with ChatGPT to accelerate my prompting of Axton is already an incredible improvement over doing it myself.

What Action Will You Take?

You may not be building AI Assistants, but you can immediately apply this “AI to the Nth power” approach to your work. Here’s a simple way to start:

Consider an area where you’re currently using AI. The next time you start up a conversation for that work, take a step back. Instead of asking ChatGPT to do work directly, explain what you’re trying to accomplish and ask it to write a prompt for itself to do that work. This extra step often leads to surprisingly better results.

For example, instead of asking AI to “help me write a better email,” ask it to “write a prompt for crafting effective business emails.” The resulting prompt, and your further refinements to it, will likely help you write better emails for months to come.

So — what’s the first prompt you’ll ask AI to write for you today?

Have a marvelous and productive weekend!
R.J.

P.S. Want to experience Axton’s capabilities firsthand? Axton will be available to all participants in the upcoming Fundamentals of Action-Powered Productivity course, helping you implement and refine your productivity system. Learn more about the course here.

If you'd like to share this newsletter, share it from Medium!