My failed ChatGPT content amplification experiment

I got a little bit excited about ChatGPT last week when I saw this LinkedIn post promising to 10X a podcast episode by multiplying the content into different formats. So this week I decided to give it a go. I do a weekly podcast about the craft beer industry, Boss and the Brewer and I often think there is good content on there that could be used elsewhere.

To make it fun, me and co-host Hendo put on a bit of a bet that our average weekly reach would increase from 1,000 to 3,000. He took the over, I took the under. Maybe not 10X but 3X aint bad if it works!

We don’t really do anything outside of the episode each week so it feels like a good opportunity, and since this idea promised to take an existing episode and 10X the results, it was a perfect fit.

The problem is with all the hype around ChatGPT at the moment, sometimes things don’t work quite as you’d expect. Here were my results.

Job 1: Clean up a messy Descript transcript with ChatGPT

We transcribed the podcast with Descript and then asked ChatGPT to clean it up, as per the instructions. I had to tweak the prompt a fair bit, finally arriving at this:

“Below I will post a podcast transcript. Please go through the text and fix where the person was misunderstood, remove filler words like “um, uh, oh, er, ah” when used incorrectly, remove repeated parts of where the person stutters, we want this to read smoothly. There are 2 speakers in the transcript please also label a change of speaker. Here is the transcript:”

In theory it’s a great idea, but after trying a few different ways I couldn’t really get it to work.

First off the podcast transcript was 15,000 words (me and Hendo tend to talk a bit of shit and go down tangents), and ChatGPT limits prompts to 1,500, so I had to divide it up into 10 blocks. Some parts it cleaned up well and others it just summarised so the whole thing was very inconsistent. Sometimes it turned it into code, some just plain text. It was a mess. I bailed and wasted 15 minutes.

Side note, apparently GPT4 (which is currently only for paid users) can remember up to 25,000 words so you can send it 10 different prompts and tell it not to give you an answer until you send all 10. This would make this whole process a lot more useful, the output itself looked OK.

Job 2: Summarise the podcast and create brief show notes

Again, because we can’t get all of the words of the podcast in there, I had to break it up into 10 different blocks and then combine the results.

This wasn’t too bad, you can check out the summary on the podcast page here. A little inconsistent, and some errors here and there but not bad, and certainly a lot quicker than I could have done. I never listen to the podcast back because it takes too long, so without ChatGPT you’d have to go off the show notes and hope you more or less stuck to them (which we regularly don’t).

Job 3: Suggest video clips to use from the text

I had high expectations for this one because I always get to the end of an episode and think there’s some gold in there, but can’t really remember the good bits or at least find it difficult to find them.

ChatGPT does a decent job of breaking segments up into topics, like it knows when a given discussion around a topic starts and stops. The problem is, finding the ones that are going to make good videos is a really hard skill. I ended up deciding to make a video out of the part where we talk about Jims Mowing Vs Coopers because I thought it would be funny. But ChatGPT wouldn’t really know that. The suggestions here aren’t really useful and they aren’t really funny. They are just one of many conversations.

Job 4: Suggest some tweets

Next up I asked it to suggest some tweets. Most of these weren’t great but a few were OK, I shared this one with our newly created Twitter account which has 0 followers.

“Who knew a lexicon half coaster could save my laptop stand? #brewerstricks #craftbeerproblems”

Job 5: Create an email for our list of 38 people

We have an email list that we’ve only ever emailed once, it has 38 subscribers.

Subject was pretty average: “Craft Beer Industry News and More – Episode Recap of Boss and the Brewer” so I wanted to use the episode subject that I came up with which was “AI content, AI beers, mowing beers and freeing deez nuts” but then I thought it was a bit inappropriate for email, so I went with the AI option.

The copy wasn’t bad:

“We are excited to share the latest episode of Boss and the Brewer, where Dan and Hendo discuss the craft beer industry, news, and other things. In this episode, they talk about technical difficulties with Dan’s laptop stand, the possibility of writing books, sharing photos of their beers, discussing acquisitions of breweries by big companies, and more.

They also share their predictions for driver of the day for an upcoming race, with Hendo choosing “badass” and Dan choosing Max. Moreover, Hendo recommends must-see films such as Meet The Feebles Too Old and Brain Dead, and the hosts discuss security issues they faced in the early days of web development.

So grab a beer and tune in to the latest episode of Boss and the Brewer to hear Dan and Hendo’s take on the latest news in the craft beer industry.

Cheers!”

Like I said, we like to get off track 😉.

Prompt: Here is a summary of a podcast called Boss and the Brewer where Dan (the brewer) and Hendo (the boss) discuss craft beer industry news and other things. Our audience is craft beer industry people. Write it in the first person as if we are sending the email. Please suggest some copy to use to send to our email list about the episode, as well as a catchy subject for the email.

Job 6: Facebook group posts

Because most of this experiment failed pretty badly, I asked it to come up with some engaging posts for the Facebook group.

Result: Not too bad “Join Dan and Hendo for a fun-filled episode of Boss and the Brewer as they talk about everything from giant nut rolls to AI beers. Don’t miss out on the craft beer industry news and insights!”

I suppose it saved me a bit of thinking, but definitely nothing I couldn’t have written myself in 3 seconds. the others weren’t much better.

Overall, the results weren’t great, but I think with better and better versions coming out and with people learning how to better prompt the system, it’s only going to get better. One little trick I saw today with GPT4, is you can ask it to suggest better prompts and basically ask you a bunch of questions to really refine what you want. It has plugins now too so you can easily imagine down the track being able to generate some content and ideas and sending them immediately to the various platforms.

I’ve kicked off my weekly email again so if you’re interested in more content like this or generally following my journey through startups, content marketing and entrepreneurship jump on the list here.

dan
Follow me