How I Lost My Weekend Working With Claude
“You’re right, that’s on me.” This may perhaps be my favorite Claude AI response. Come on, don’t pretend you haven’t gotten this back from a prompt.
And by the way, who’s the Anthropic AI developer that came up with this as the go-to mea culpa every time Claude directs us down a rabbit hole that compels us to prompt: “No, that didn’t work, look at this screen shot! Do you even know what you’re doing?!”
Let’s face it, while AI opens pathways to do all kinds of stuff that we couldn’t imagine doing, it also sucks us into black holes more often than we’d like to admit. I’m here to admit it.
Allow me to share how I spent a recent weekend. While I grant you that I do indeed need some new hobbies (just don’t tell me golf should be one of them), I do find tinkering with AI to do all the things I could never do on my own kind of like a hobby. Yes, a weird hobby I know.
The saga began Saturday before last, mid-morning. I picked up the old laptop, hopped on the couch, and spun through ESPN+ to find a college baseball game to put on in the background. It was time to start building out the updated version of my digital marketing stack to support my nonprofit MyNextPlay.org.
The plan was pretty simple: export contacts and their accompanying organizations, or “Company Name” (more on this later), from my donor fundraising focused customer relationship management (CRM) tool Givebutter to Hubspot.
Of course, before starting this endeavor I iterated within a Claude project to confirm which new CRM could streamline my nonprofit’s go-to-market efforts. When I followed up in the same project conversation to ask Claude how to go about actually executing this transition, he was so confident. He mapped out a plan in a very specific number of steps, assured me this would all be done in less than an hour, and sent me on my way to start exporting data from Givebutter. (Side note: I’ll be referring to Claude as he/him even though to be honest I’ve never seen Claude post preferred pronouns?)
Let the Gauntlet Begin
A few hours after updating my exported list of contacts from Givebutter, and then reformatting and cleaning the data with the proper field names, I got the first warning sign that trouble was brewing. Yep, the first import to Hubspot resulted in almost double the number of contacts that were in my “master” spreadsheet.
Allow me to say the quiet part out loud – Claude doesn’t know how Hubspot works. Instead of figuring out how to get the single CSV import to deduplicate the “Company” objects correctly, Claude suggested breaking the spreadsheet into two .xlsx files to sequentially import the unique list of Company items first, and then follow that by importing the Contact records.
Did I mention above that I need some new hobbies? Okay, back to the story.
With a new iteration, the dual spreadsheet approach seemed to deal with the duplication issue, but then good old Claude failed to alert me to the other nuance of Hubspot: Company Name isn’t the unique identifier that it had assured me would be the secret “key” to binding data from the two sheets. And to make matters worse, Claude hadn’t recommended turning off the annoying Hubspot feature that looks at contact email addresses and then assigns (often erroneously) company names.
Oh for Christ’s sake, “What is going on here?” I thought to myself. Actually, I think I blurted it out loud, not just because Claude was clearly dragging me into that special world of “AI merry-go-round,” but because it was now almost dark outside. By the way, it doesn’t get dark these days until well past 8 p.m.
For the rest of the evening it was cycle after cycle of me trying something and then uploading a screenshot to inform Claude what I was “actually” seeing, not what it thought I should be seeing. By the time I called it a night, my laptop’s desktop screen was so littered with screenshots I could no longer see even one pixel of the lovely wallpaper image Apple added during my last operating system update.
Sunday Morning Solution
During my morning walk Sunday I couldn’t stop processing what to do next in my dance with Claude and Hubspot. I think we ended Saturday night with an agreement that we’d go back to trying the single CSV upload option using something called multiple objects importing, an idea that actually came about when I ran out of tokens with Claude and decided to ask Gemini what the heck to do.
While on my stroll through the Golden Gate Park panhandle, I was asking Claude (via the mobile app) for an updated plan that would ensure this multiple object import strategy would work. I hustled back to my home office, clearly waving the “I have no life” white flag as I fired up my laptop and started back in on “Project Hubspot”. Two hours later came the best moment of the weekend: Claude recommended that I go with Brevo, not Hubspot.
“Nope!” I told Claude that was an absolutely unacceptable option, and proceeded to furiously type something along the lines of “You’re wrong and you gotta figure this out,” and that’s when I got the infamous “You’re right, that’s on me” response from him.
Finally, after another hour we got to the bottom of things. Just to bore you some more about Hubspot, apparently “Company Name” can’t be assigned to contacts without a unique Company Domain Name identifier, even if the domain URL is fake and doesn’t actually go to a live website. And, as it turns out, for many organizations in my contacts you can’t use their actual URL because Hubspot doesn’t accept the format of their website URL.
Again, for Christ’s sake. It’s 2026 and these cloud-based super SaaS software platforms still can’t figure this stuff out? No wonder the SaaS apocalypse is upon us. I’ll spare you the rest of this Claude vs. Hubspot drama and tell you that after some finessing with dummy URLs I got what appears to be all the unique contacts imported, with the right number of unique companies, and properly showing the Company Name for each contact (even if the domain name for a large chunk of contacts isn’t real).
Learnings?
What did I actually take away from spending basically an entire weekend trying to get AI (Claude) to help me set up a supposedly important software platform (Hubspot)?
Don’t ever spend your weekend on an inane task like this. Even playing golf would be less painful.
Assume that the AI doesn’t really know how to do the very specific things you need it to do, and as a corollary, don’t expect that you’re going to be able to craft prompt instructions so detailed that the AI will get your task completed correctly.
Expect errors, frustration, and needing to re-do things over and over.
Don’t get too frustrated, don’t give up, and don’t let the AI give up on finding a solution.
Find a consultant on Upwork, Fiverr, or through your network to help you actually do the task the right way, for a fraction of the time. Whatever you pay them is more than worth it.
I will say, despite spending way more time than I should have, I got the project completed (sort of), figured out a fair amount about how to navigate Hubspot, and learned even more about how to navigate the nuanced dynamics of how to utilize AI to get more done than I realistically could have without it.
But still, I need some new hobbies. Any ideas for me?
NextPlay>Forward AI Disclaimer: I very actively use artificial intelligence and large language models to generate the content you read here, but I do review it and edit it to make sure it can be generally useful to people who read it. Keep in mind that AI can make mistakes - check important information. Let me know if I make any errors and I will correct them.



heya - yah, part of the challenge is there's a new tool and/or way to pull AI into your workstream, so it becomes a major time investment just getting up the learning curve to get going. then there's the iterations based on fact no matter how much context and description you provide, it doesn't deliver exactly the product you envision.
So relatable... I started a personal finance tool yesterday morning, and poof it was 9pm. Trying to remind myself this is the awkward adolescence like dial-up. Have you tried Claude for Chrome? I wonder if it could handle this on its own and verify its work... maybe too early but soon.