Day 2 of AI - IG posts
Claude
I started two days ago in ChatGPT. I switched yesterday to Claude. First impressions — Claude was superior in coding and in responsiveness over time. ChatGPT tends to get slow as the conversation goes on. Claude is only slow when you ask it to perform big tasks. Claude feels more “right".
adsbot
First, updates on adsbot, the effort I reported on yesterday. I still have not been granted LwA access from Amazon, which means I cannot pull ad spent data. So the engine is half-baked with just sales data. That said, Claude and I worked out the sales data side of things. We had one interesting find—the data on the Amazon Dashboard for Royalties is wrong. It ads British pounds to dollars without converting. I learned I’m actually making about $10 more per week than I thought I was.
I also pushed on the Claude Cowork with my ads spend data (pulled directly since I cannot get LwA permissions yet). I was a little disappointed with Claude’s response. It told me to kill some ads b/c they are low performing. I pointed out that Amazon ad spend data is quite poorly correlated to sales. It then agreed with me and said that 87% of my sales are unaccounted for in the Amazon ad spend data. I replied with something like, “no kidding, that’s why I only use it for positive correlations. When I do have ad sales data, I chase it.” And it told me that was a wise approach. I was hoping for more insight. I’ve had a streak of high spend with no sales this week, and I poked on that data with Claude Cowork, and it never said anything I hadn’t already realized for myself, and its advice boiled down to, “let it ride and see what happens”. So, kind of disappointing.
igbot
I started a new effort called igbot. I wanted something that would create IG/FB posts for me based on my actual book content. Before I describe details, let me give an example of the results:
https://www.instagram.com/steven.j.morris.writer/p/DVIJq5KkuTw/
I was impressed. Some technical details:
It uses AWS to store pics in S3 buckets, while the post text gets stashed in DynamoDB, before pushing to IG/FB. It fires through Lambda and scheduled via EventBridge. Keys are stashed in Secrets Maganer.
It uses the Meta Developer API to actually push the data.
It has tracking capabilities (as yet untested) that we will use in a feedback loop to lean into what works.
It calls Claude’s Anthropic API to generate each post based on summaries of the books and characters (okay, it didn’t do a GREAT job of creating the summaries… it would swap which books things happened in and get details wrong. I had to intervene.)
It can store types of content to create. For example, I have a book release coming up on April 13. Starting March 16, it will sprinkle in (25%) posts about the new release. And after the release it will sprinkle in posts reminding people to review (which maybe I won’t do, but it is an option).
I’m storing all of the code for both of these bots (if they really are bots) in github. I did not capture all the deployment steps, but I think I will. Maybe there’s people who want to do something similar for, I don’t know, getaway rentals in Fredericksburg.
Right now, I have to run all of the generation from my laptop. Or a laptop where I can check out the code. I’ll think about whether it’s important to be able to do it from somewhere else… say web based. That might be nice. Then I could keep posting while living it up in my vacation rental in Fredericksburg.
But I can run multiple post generations and store them for deployment. Though now that I think about it, I didn’t create a scheduler, so I still have to press a button to trigger them. Feature request!
Broader View
Contextually, I got a bid from a book IG/FB marketer. $300 a month to do this same thing. And I honestly think this post from Claude is better than what that team generated. Claude is $200 a year. There’s lots of other bits that cost… the Agentics API, the AWS eventually charges you, and I pay for midjourney for the images. The Agentics & AWS are dollars per year at the rate I use them. Midjourney is the most expensive bit—maybe I will phase that out for a cheaper option. But midjourney most consistently matches my asthetic, so maybe not. So this is definitely cheaper, and I understand how to tweak the input for the Agentics to lean into whatever we find works.
Next step is posting for a while and then testing the feedback loop. Honestly, I’ve lost my audience on IG/FB, so it will likely take some time to get momentum. And I’m not sure how much IG/FB matters for book sales. But this is a fairly inexpensive and non-stressful way to find out!
