In this episode, eight months of newsletter production using Claude with custom instructions, project knowledge, and exact transcript inputs still requires 1-2 hours of line-by-line rewriting per issue. Despite 55-65% open rates proving the final quality, AI-generated drafts consistently miss the authentic voice even with growing training data—revealing high-quality writing remains firmly protected from automation.
Key Takeaways:
- AI Struggles Despite Perfect Training Data Feeding Claude full podcast transcripts in your exact speaking voice, comprehensive project knowledge, and dozens of final newsletter examples still produces drafts requiring massive rewrites—the tone and personality consistently fall flat.
- Custom Instructions Bloat Without Quality Improvement Iteratively expanding custom instructions to hundreds of words didn’t improve output—sometimes stripping them down and relying purely on project knowledge performed better, suggesting AI can’t yet extract signal from noise effectively.
- Newsletter Writing as Competitive Moat The inability of the best available AI model to replicate authentic voice after months of training means skilled writing remains a durable skill—those who can convey thoughts into compelling words maintain significant entrepreneurial advantage.
Listen to the full episode for the honest assessment of AI’s current limitations in content creation and why writing skill continues to compound as an irreplaceable asset.
(Podcast transcript below)
Welcome to Get Serious. Today, just wanted to share a little bit more behind the scenes, specifically regarding AI use and its capabilities and how I think that high quality writing is just not in any danger of being replaced at the moment for my own sake.
I don’t know, I kind of go back and forth with it in terms of like, yeah, if you get to snap your fingers and just have better quality copy being available, would I take that versus, you know, kind of preserving the skill set of being a higher quality writer to still, you know, justify additional value into the marketplace. You know, there’s a push pull there.
specifically in relation to emails in our case here. Well, there’s a few different avenues. mean, there’s just general email drafting. And I know I’ve commented on like Sarafey.
and in practice, honestly, I find myself utilizing the pre-drafted responses fairly infrequently, honestly, like less than 10 % of the time. Usually for short emails, it might be a little bit better. For longer ones, I might get a few ideas on it, but it also can hallucinate certain responses and just come up with numbers related to like…
funding decisions and so forth that just aren’t accurate. So that’s still a work in progress. out of any work tool, that to me might be like the holy grail is, yeah, if I could just have an auto communication mechanism, both for text messages and emails, ideally for phone calls too.
but I haven’t seen anything that quite hits the mark, especially for things that require more nuanced responses or even like medium level of nuance. It just has not hit the mark yet. But then on like, know, actual writing projects, like for instance, you know, the Serious News newsletter and now that, you know, been doing more email marketing for Lampricer as well.
So usually for serious news, again, weekly newsletter, again, a little bit more on how the sausage is made. And I might have talked about this before, but just a rehash on where we’re at, especially in the production side and how AI comes into play here. So I have a separate Claude project that has custom instructions related to my overall tone and style, general breakdown.
how those emails work and they follow a relatively similar framework and tone of voice for the most part. And so I’ve tried to just constantly reiterating with Claude on how to update the custom instructions. Plus every time I write an email, the final copy will get added to the project knowledge so that it can pull, okay, precisely.
what have some of these past emails look like. So then you can just update it. like most of the time, my content from or for the newsletter has been covered within a previous Get Serious podcast episode, whether it’s one episode, two episodes, three episodes, there might be a little bit of a mix. And so usually I’ll throw in…
you know, I’ll go into that cloud project and I’ll throw in a copy of the transcript from the podcast and then I’ll throw in, you know, maybe a little bit of directives on how I want to structure the overall email, which, you know, core points are more important here than again, like, okay, pull in, you know, maybe it’s most related to this particular previous email in terms of the overall style and that.
Yeah, just wait for clog to populate. Plus I’ll ask it to give me some anticipated high performing subject lines to bump open rates. Like open rates are pretty solid. Actually, I don’t know what industry averages even are, but like I’m routinely between like 55 to 65%, which I think is really high for open rates. So kudos to the audience for wanting to.
to read what I have to write. We’re always looking to improve that as well. The click-through rates usually no higher than 5-ish percent or so, which is probably still pretty high on that.
as well but I know there can always be further improvements from that regard. the unsubscribes are fairly low to like less than a percent at the time. I’m not sure. I mean usually we’re kind of adding, anybody we have contact with will be added to the list and they can always unsubscribe if they want but I just attribute it to.
I’m churn or they don’t like my tone of voice whatever Or it’s just not as relevant to them. Okay, that’s fine I don’t want to be delivering emails to You know folks who aren’t interested that’s gonna hurt my metrics overall so you want to continue to warm warm your correct audience over time is always the goal here plus I don’t want to pay for People who are just inactive on my list And
So that kind of goes back to how I’ll orient the Claude project. And I’ve been doing these weekly emails now for…
at least four months roughly, somewhere around there. And I think only once I had a email that popped out that was like close or like cut copy from clog that popped out that was like 80, 85 % done and required fairly minimal edits. The rest of this time, especially lately, honestly, I felt like it’s just been a bit.
worse or maybe it’s like just my eye for the writing has become even keener. And like the Claude models have improved too. Now it’s on the Claude 4 Sonnet. Just got better at copywriting and Claude is generally known as being best for content production out of all the LLM models. Plus again, like my project knowledge continues to grow. And like the voice just oftentimes it’s not there and I’ve had to reorient the custom instructions.
Like, okay, anytime, hey, here’s the final copy. Like, you still missed a lot. How should we update the custom instructions? Like, sometimes it’s relevant and like the custom instructions got so bloated to where I felt it was just off and I’m like, okay, let’s just like boil it down to be less restrictive and just utilize the project knowledge more. And that hasn’t worked quite as well either. And again, like I have all my core voice.
which I think is just a…
It’s still really hard for AI to pick out the signal from the noise here. The signal should be super strong. I’m giving it podcast transcript, like my exact phrasing, tone of voice, style, how I’m conveying a subject. All of that is right there, the way that I’m talking. Plus I have now a growing number of…
exact final copy for previous newsletters following a certain style and trying to update the custom instructions and it still can’t routinely nail the copy or even like really get it fully there. So usually I’ll just take that initial copy and I might try a couple iterations like, hey, you got to try hard. Like this just sounds like AI slop. It’s way too bland. It’ll come back. Try to match my voice a little bit better. And then generally I will
copy and paste it into MailChimp, which is the service I’ll use to send out the email. But then I’ll just go line by line, just continue to edit, continue to edit. That’s usually my writing process. Is, okay, at least I have a draft to work with, and then I’ll just adjust, adjust, adjust, do I need to move things here, et cetera, to get it to my voice, and then I’ll go back to Claude, like, okay, here’s what I really like here. And then I’ll also ask it, like, hey, anything else that you might suggest improving here?
and it can provide feedback in terms of was your hook good enough for any other adjustments that I want to consider. So it can also serve as a copy editor, which its functions in that regard is definitely better than its ability to produce copy, which no surprise, because producing the actual copy is just, it’s a more difficult skill. But like in terms of time for me to do this, I usually ranges an hour, hour and a half, sometimes two hours plus.
to really nail these. Again, good open rates, good readership. I continue to get praise from and there’s good reviews from readers on this particular output. Plus, again, we’re syndicating and every couple weeks we’ll have an article go live on Ari Tipster as well, which has a lot more eyeballs.
So I put a lot of effort into doing this right. But again, I think this is just an example where it’s like even having all these advantages, my exact voice, transcripts, existing knowledge and like, you know, utilizing the best AI model currently available on the market still misses the mark to where I’m having to do an hour or two hours of just adjustments and rewriting.
to get it to hit my tone of voice because I know I’ll just lose the readership and I’ll feel bad, personally guilty, hey, if I’m just pushing AI slop out there, that’s not matching my true intentions. Because the way that I’m talking to you now via podcast, that is a tone of voice that should be conveyed over into the serious news newsletter. And again, I think I’m doing a good job in regard to that.
based on the response. But like that’s always the one to one goal there because you want the true personality and intentions to show up. So all of this again is to say that, I mean, these AI models are always getting more powerful, right? But to like…
I mean that would make my job a lot easier, Especially again, if I have these, like if I can throw a transcript in, okay, just reorient this to the structure of the email, boom, can just post it and I’m done in like 10 minutes to get it out. That would make my days, my weekends a lot easier. But like we are still very far away and I’m not even writing that long of article. I mean usually less than thousand words, like 500 to maybe 750 words roughly. Like it’s not much content.
So imagine if you’re doing much longer form content or even books, etc. All of a sudden that complexity is just through the roof in terms of what AI could potentially produce. And I hear this from other newsletter writers or full authors as well here. again, we’re just a while off from…
Unless things just dramatically improve with some of these upcoming models, which I’m sure they will eventually get there. It’s just not Not really that close as of yet, especially when you consider the you know additional level of complexity to consider so in the meantime Still a valuable skill set and you know if you can convey your thoughts into words I mean if you can be a storyteller like
I don’t know, is there more powerful skill to have as an entrepreneur? It’s up there. Certainly within the top five, you can make an argument for the number one skill. in the meantime, it continues to refine my writing skills, just keep getting better and better and better. Endlessly, because I’m not stopping anytime soon.
And I’ve been recognized for writing before, like, even growing up in school, et cetera. I took a lot of pride in writing and was good at it, right? And so now that I can show it more in a consistent professional manner outside of just general emails and so forth, transactional emails internal to your Gmail system or whatever, which is a skill in and of itself, but a whole newsletter set up.
I appreciate the ability to produce and provide value for people to actually read and consume. It does make me feel good. A lot of times, I might procrastinate and just be a bit annoyed. You’d have to get started then sometimes in the process, in which the AI was just better at matching my voice, so there was just less rewriting going on. But every time that I reach a final…
product, like I do feel satisfied and proud to send it out. And as soon as that doesn’t happen and I try to rush it, I know the response will be very negative from that, both from myself and from the audience. So like that is my promise to you to always be delivering quality and not mailing it in. just wanted to share again some more behind the scenes and thoughts behind this process.
SeriousLand.Capital for any of your funding needs, Land Daily Diligence Facebook group for zero cost review of your land deals, and LandPricer.ai for the most reliable land pricing tool on the market. that, subscribe and share. Take care, everybody, and talk to you tomorrow. Bye.
