Day 14: Accelerating AI Content Production
It’s been a crazy week at Vacation Tracker, so Day 14 and Day 15 are being worked on closer to the end of the week.
Today we’ll be continuing to explore how CofounderGPT can help us with content and we’ll try to see how we can make the process of creating content with our AI cofounder more quick and easy. In Day 13, it was very time consuming to create the first long form article. That method of working is not sustainable long-term if we want to produce one AI-created long form article per week to help our SEO efforts.
Let’s try to figure out a faster way to work today. If we can’t, we may have to scrap writing one article per week until ChatGPT gets a little better at producing long-form articles without so much supervision.
Sending our newsletter
I realized after publishing Day 13 that I didn’t explain how the emailing part of the newsletter work. In fact I never discussed email addresses for Knowlo in general so I would like to take a moment in today’s article to explain all that.
We don’t have @knowlo.co email addresses yet. And we probably won’t for a little while because there is no reason for us to have them yet. But we do need some email address which is not Slobodan’s or my personal email address. We have an old Google Workspaces account for our first business Cloud Horizon so we decided to set up a temporary email account with the @cloudhorizon.com domain. This will let us start sending emails from Mailchimp without actually having @knowlo.co email accounts.
As a business, there are really only two serious options when it comes to email providers for your own domain: Google Workspace and Microsoft 365. And if you are not able to use anything other than Excel and Word (Microsoft editions), well then there’s really only one game in town for you.
If we did not have a Google Workspace account and we were starting from scratch, we would probably go with Google Workspace again. I’m just used to it after all these years, It’s the one I’m most familiar with. But we did switch to Microsoft 365 for Vacation Tracker and it has its advantages and disadvantages. For startup world though, I think Google Workspaces is a little simpler to get started with.
Both Google and Microsoft bundle their productivity suites in with their email services, so for me it makes no sense to go with a smaller email service provider if you have to pay for your productivity tools separately.
So with this minor technical detail out of the way, Let’s get back to creating content with CofounderGPT.
What did we learn from writing our first article?
The last article CofounderGPT wrote took much longer than expected. I need to be able to get a long form article out in 40-60 minutes maximum, once per week. And there is a mechanical process at the end which is difficult to automate right now using CofounderGPT which is actually publishing the article.
When our AI cofounder finished writing the article, there were a few steps that I needed to do to complete the article and publish it:
- Edit the text and suggest improvements with CofounderGPTs help (25 minutes)
- Make a list of possible images for the blog post and make a selection (15 minutes)
- Generate prompts for Midjourney using AIPRM prompt template (10 minutes)
- Create Midjourney images and make a selection (20 minutes)
- Transfer content from Google Docs to WordPress and publish (20 minutes)
Just this part took an hour and a half. And then there was the writing and fiddling with prompts to get a decent quality article out of CofounderGPT. If we want to write and publish one article per week, I’ll need to be able to have CofounderGPT research and write the article without much input from me. That should take 20-30 minutes and I need to compress the pre-publish list above down to 15-20 minutes if I can. I’m not sure this is possible but we’re gonna try.
New way of researching
I follow a few people that post about AI on Twitter. One of them is Rowan Cheung and he posts regular updates about What’s happening in the AI world. If you’re on Twitter and you’re interested in AI, I highly recommend you follow him. He also has a daily newsletter called The Rundown.
Anyway, he recommended a relatively simple method for getting ChatGPT-4 with Browser to do research for you. I was thinking about taking his approach and modifying it a bit in the hopes of reducing the section-by-section research and writing we were doing last time.
Here is Rowan’s suggestion:
I tried to do literally the exact same thing as Rowan. But for me, the first 3 times it completely crapped out:
Then after the fourth attempt at getting this list of 10 things that happened in AI this week, this is the result it produced:
This is far from optimal. All these results are about Google. Yes, Google had a major event this week and they made a lot of important announcements about AI. But these are most certainly not the only AI news that happened this week. I need a list of more than just Google news because we want to pick something that’s related to what we’re doing and then try to write an article about it with CofounderGPT. If you click on all the source links (the little green numbers at the end of each item) of the news it presented in the list, it takes you to the same article on CNBC. So CofounderGPT took the lazy way of doing this. It found 1 article on CNBC that summarizes the news from Google I/O, and then basically summarized that article. This is not what I had in mind.
Let’s try again with a slightly modified prompt:
Even with an improved prompt which specifically asks for news from different companies, this is the result that was produced. And it got cut-off at the 6th item as you can see in the screenshot above.
Researching with Bing
This is too time consuming already so I’m abandoning ChatGPT-4 with Browser (a.k.a. CofounderGPT with Browser) for researching articles for now. It’s too buggy and inconsistent. I’ve found that Bing does a better job of doing web research so I’m going to try the modified prompt I used in Bing instead:
This seems like a better result when you scan through it quickly. It is 10 different pieces of news from different organizations. But 9 out of 10 of the results are about researchers in various universities discovering something. All of the 9 are related to some kind of AI health-related discovery. Having spent the last 90 days being best friends with various AI tools, I know they are prone to hallucinate sometimes. This list looks very fishy and I suspect that most of these news did not happen in the last 7 days (if they even happened at all).
Clicking the first item on the list brings me to an article about Lithium (here is the article it linked to). Nowhere in this article can you see anything about Alzheimers or AI. Clicking on the second link brings me to the Artificial Intelligence section of Wired, but I can’t see this news from MIT and Harvard. In fact, a quick Google search of the first three items on Bings list shows that it completely fabricated these news. As in, it literally invented these 10 things that did not actually happen in AI this week.
I’m not even going to waste my time trying another prompt. If It’s going to hallucinate results on the first try, It’s not going to be a tool we can reliably use for research for now.
Well, It’s clear that we can’t rely on the AI tools researching current events and writing articles that tie in with what we’re doing at Knowlo. So we’re going to have to try a different approach. I guess we’ll have to write purely theoretical articles about AI and knowledge bases that do not require an internet connection.
Let’s ask CofounderGPT to come up with some ideas for the next long form article in our series:
Last week, you wrote a long-form article for our blog about “Harnessing the Power of AI for Niche Industries: Uncovering Hidden Gems.”
Today, I need you to write another long-form article as part of our AI-generated content series. You, CofounderGPT, will be the author of this article. Since we are building Knowlo, an AI powered tooltip product, I’d like to write about something related to how artificial intelligence can be used for tooltips, knowledge bases or help desks.
Our audience for this article are product managers and CTOs in software businesses. Keep this information in your mind, and do not explicitly state it in the article please. I am telling you so you adapt your writing style and content for this audience.
The length of the article should be about 2,000 words divided across multiple sections.
Give me 5 possible subjects for this article. Please don’t go with something super obvious. Be creative and come up with something interesting and engaging for our target audience.
Even though I explicitly asked for non-obvious ideas, I got 5 very obvious ideas as a first result. So I told CofounderGPT these ideas are too obvious and then it produced 5 slightly less obvious ones:
I kind of like the last subject, although I’m not sure how CofounderGPT is going to write about this without doing an internet research. So we’ll change the title slightly to “How AI will Shape the Next Generation of Self-Service Platforms” and that will let us talk about the possibilities rather than What’s currently happening.
Preparing an outline
Next we’ll ask CofounderGPT to do a quick outline of the article:
Ok let’s go with #5. Remember, our audience are product managers and CTOs of software companies. Remember, we want the article to be 2000 words or more. You are the author of this article and an expert on this subject. Your job is to come up with content which will be appealing for our target audience and which they will want to share with their peers. Please come up with an outline for the article.
I’m not going to overthink the outline this time, we’ll go with the first suggestion that CofounderGPT comes up with:
This outline is pretty decent actually. So I asked CofounderGPT to expand a little bit on each section and to include some more ideas of what we could write about and a suggested break down of the word count. Two prompts later, I had an outline which we could work with.
There is a reason why I want a clear and detailed outline. It’s going to be part of the prompt template that we are going to use to create articles from now on. If it works, we’ll use this template every time we need to write an article, we’ll just substitute the outline and keyword.
I would like you to act as an expert SEO writer to write a long form blog post for our Knowlo blog.
Title: Customer Support, Rebooted: How AI will Shape the Next Generation of Self-Service Platforms
Blog outline:<INSERT OUTLINE>
Target audience: Product managers and CTOs at software companies
SEO keyword: knowledge base in AI
– Write a long form blog post with at least 2000 words and write it as specified in “Blog Outline”.
– Write each section in the suggested word count
– Write the article in a way that will be engaging for our “Target audience” and which will be optimized for the “SEO keyword”.
This prompt produced a 982 word article. Last time we went section-by-section and ended up with 3,800 words instead of 2,000. For some reason, CofounderGPT is not good with understanding how to follow the word counts. But at least the article came out ok in terms of quality so I’ll move to the next step.
To complete the article, we need a cover image and another image somewhere in the middle of the post to make it a little more visually interesting. I asked CofounderGPT to come up with an idea:
Please suggest two images for this article:
1) a cover image
2) an image for somewhere in the middle of the article
Both images should be related to the content.
The ideas that it came up with are “a blend of interconnected gears or puzzle pieces, symbolizing the integration of AI technology into the support process for the cover image”, and “an artistic visualization of data patterns, where abstract shapes and lines are arranged to represent the data analysis process”. I asked CofounderGPT to write a descriptive sentence for each idea and put that into Midjourney. I didn’t want to waste a lot of time so I picked one of the 4 options from the first batch of suggestions Midjourney gave me:
Publishing and Summary
This time I skipped editing the article and getting suggestions on how to improve it. And because of that, I skipped the whole Google Docs step and pasted the article directly into WordPress. I also didn’t go fancy with the Midjourney prompts, I just selected the two sentences above from the description CofounderGPT provided and ran that through Midjourney.
All in all, writing the article took more than 40 minutes, but I do think if I stick with this template and try different versions of the prompt, I can get the article writing with CofounderGPT down to under 40 minutes. The pre-publishing process took 20 minutes this time, which is in line with what I wanted.
In both cases, I had to reduce quality standards in order to get it done so quickly.
Conclusion about AI-generated content
I’ve been using AI tools obsessively over the last two months. They have made me more productive and I see a crazy amount of potential for what AI can do to literally make people superhuman. Having said that, there is a reason why most of OpenAI’s latest products have “Alpha” labels on them.
At a high level, when you’re building software, you generally start with a prototype, then you evolve that to an “alpha” version, then a “beta” version and then finally a release version. At each stage you let in more customers and they give you feedback to improve it for the next iteration.
Anyway, if you want to produce a high quality long form article using ChatGPT that’s SEO optimized, the whole process from writing to publishing will still take you several hours. Unless you want to sacrifice quality standards the way I did with today’s article. that’s how it is today and I plan on writing a separate article about the reality check I’ve had over this recently.
But things are moving very fast in AI right now and it might be that within a few weeks you will be able to write and publish a super high quality article with ChatGPT in less than an hour. Next year at this time, I’m 100% sure you will be able to do it.
For Knowlo, we’ll do some AI-generated content but not on a regular basis like we planned initially.
No money was spent today, just a few hours with CofounderGPT.
Time spent today: 4h
Total time spent: 91h
Investment today: $0
Total investment: $913.54 USD
Paying customers: 0
Slobodan is going to provide an update on his progress building the prototype with CofounderGPT. I’ll be working with our AI cofounder on adding some more content to our website so we can run a small ad campaign to try to get some beta customers.