How ChatGPT cut the time I spent setting up this newsletter by 50%
A step-by-step example of how AI can save your team hours each week
If I had to write the content for this Substack entirely from scratch myself, every time, I probably wouldn’t be doing it. I’m trying to launch a consulting business from scratch, I need to generate revenue, and content takes a lot of time to produce.
Well, content USED to take a lot of time to produce.
Now, with ChatGPT’s help, content takes me about half as long to produce as it used to. In addition, I generally don’t feel any need to try to wrangle someone else to give me feedback or proofread.
To launch this Substack, I needed the following chunks of text:
Short description
About page
Welcome email
First newsletter
Let me walk you through how I generated the first three of those in a fraction of the time it would have taken if I’d been writing them from scratch on my own.
Note: This is an context-setting post on GPT-4 capabilities for readers who haven’t played around with the most recent LLMs much yet. For those of you who already use ChatGPT regularly yourselves, you might find today’s content to be old hat. Stay tuned to future issues for more in-depth conversations!
Tool selection
The first question is always: What is the right tool for the task I’m performing?
GPT-4 by OpenAI is the most powerful LLM available for consumer use right now. It is rivaled only by two other options: Claude+ (from Anthropic), and Bing’s search engine. Bing’s search engine allegedly runs on GPT-4, and has the advantage of having access to data post-2021 (when GPT-4’s training data ends). But:
I’ve sometimes found that GPT-4’s answers are better than Bing’s (no real idea why), and
Using GPT-4 directly through OpenAI’s interface allows me to save the conversations — and continue them even if I’m disconnected in the middle — whereas Bing doesn’t.
So, I use GPT-4 through OpenAI’s ChatGPT Plus interface as my default, and switch to Bing if I’m doing something that really needs real-time web access.
ChatGPT access costs me $20/month; you can sign up here. Note: Every time you start a new chat, make sure you select GPT-4 manually from the dropdown menu:
Background information and the ask
Now, ChatGPT does not have access to the internet. GPT 4.0 knows nothing about anything after Sept 2021, including what my company is doing, the capabilities of recently released AI products, and the recent public discourse about generative AI.
So, I began by giving it a bit of background information about AI Impact Lab and my goals for the newsletter. Then, I ask it to draft the three pieces of evergreen content (the Short Description, About Page, and Welcome Email):
In less than a minute, it comes back with;
Pretty credible, right? But not perfect…
The iterative process
Now, I can’t just copy, paste, and hit publish. I need to read all of GPT’s content very carefully. Partially because of the potential for AI hallucination, and partially because ChatGPT doesn’t actually have very much context — there’s a lot I know about the business and the types of stakeholders that I haven’t told it.
After reading and thinking for a minute, I have a concern: The content is too unreservedly positive about AI's potential impact on the world. While I am excited about the upsides, I need to acknowledge that there are lots of risks and potential downsides as well.
I’m tempted to just start rewriting the content myself, but I quickly realize that’s the OLD way of doing things. Instead, I prompt GPT:
And it does! I’ve circled some of the key changes it made:
As you can see, it added phrases like “responsible approach to AI” and entire bullet points like “Strategies to address potential risks and challenges associated with AI adoption.”
At this point, I copy-pasted the text into the relevant Substack fields, making only minor edits.
I don’t want to oversell GPT-4: I’ve found that this “about page” type of static/evergreen text is one of its sweet spots. For instance, while it was helpful in cowriting the first newsletter post, it might have saved me only 30% of the time rather than the 70%+ that I estimate it saved me on drafting these evergreen fields. But that is still very significant — and the tools are likely to keep getting better and better very quickly.
Can you imagine a world where every piece of content your entire team writes, whether for internal or external consumption, takes only about half the time it does now? How many hours a week could you work less? What impact could you have with the time you saved?
Prompt engineering
The skill of successfully getting a generative AI tool to produce the content you want is called “prompt engineering.” Prompt engineering will be one of the most valuable skills you and your employees can gain in 2023. You don’t need to take it from me:
Entrepreneur: What is Prompt Engineering, the hot new tech job in AI?
Business Insider: AI 'prompt engineer' jobs can pay up to $335,000 a year and don't always require a background in tech
The examples I showed above — giving background information and asking for modifications to earlier responses — are very basic forms of prompt engineering; I’ll delve into more advanced methods in future posts.
The good news is that prompt engineering is an eminently learnable skill — and I’d love to help you train your team on it! If you’re interested, check out my offerings and contact me through www.aiimpactlab.com. And of course, if you’re not already subscribed to this newsletter, now’s the time:
This example of cowriting with ChatGPT is just the tip of the iceberg when it comes to the incredible potential of AI tools for content creation, especially when used thoughtfully and iteratively. As we continue to explore the world of AI together through AI For Good, I hope this example provides some inspiration for how you might integrate AI tools into your own work to maximize your organization's impact!
I noticed when you prompted for the rewrite you used the term “gung ho.” What do you find are effective and ineffective ways to use informal language like that in prompts? Does using less formal language in the prompt result in less formal language being generated?