Hearts 'n Tales, my podcast, lives on a server I control. I self-host it. No middleman, no third-party media host, just Apache and me. That means when I want real download stats, IAB-compliant numbers, not vanity metrics, I can't just install a plugin. I have a problem to solve.

So I solved it. Over one weekend.

Here's how it actually went, because the "how" is the interesting part.

Saturday morning:
I dumped my thoughts into an idea.md file. Nothing fancy. Just me thinking out loud in text about what this thing needed to do and why. Then I turned two AI agents loose on it, my ARCHITECT and my DBA, and told them to argue about the design. I limited them to 4 iterations so they don't go nuts and burn tokens like they're a pile of leaves in the fall.

I went out and trimmed the hedges and watched 2 really bad movies because The Lovely and Talented Kathy was out, so I get to watch the movies she won't.

When I was done, they had completed loops 1 through 4.

Verdict: 8 load bearing blockers. Not "nice to haves." Things that would break the system if I ignored them.

Saturday afternoon:
I kicked off loops 5 through 8.

Verdict: Down to 6 blockers between them.

Sunday morning:
Before church, I kicked off loops 9 through 12. I came home to a green light from my DBA and an Architect still holding out for 3 more fixes.

Sunday afternoon:
Before I went car shopping with my son (he bought a used Tesla) I kicked off loops 13 through 16. When we got home I was pleasantly surprised with the verdict.

Verdict: Both agents green. Design approved.

Sunday evening:
Before we started a movie, I turned the applicable agents loose to actually build the thing from design.md.

Monday evening:
I deployed. It wasn't exactly smooth. It took a couple of hours to work through an issue the Architect didn't anticipate. To be fair though, I did not specify the deployment environment. Had I done that, it probably would have been smoother.

We also found one other non-trivial issue once traffic started hitting the system. That took another 30 minutes to fix.

That's it. That's the whole project. I checked it this morning and I can see what people are listening to and where the traffic is coming from.

I've got a stats tracking system built specifically for my podcast, logging downloads the right way, off Apache logs instead of a redirect prefix, because a redirect prefix means adding a hop between a listener and their audio file, and I will not slow down or risk breaking audio delivery for the sake of a metric. That rule was non-negotiable before I wrote a line of code.

Now here's the part I actually want to talk about.

If I had gone shopping for this instead of building it, I would be paying somewhere between $10 and $20 a month for a podcast analytics SaaS. Forever. That is not a knock on those tools. Some of them are good. But I didn't build a SaaS product here. I built a bespoke tool designed specifically for my podcast, shaped exactly to my server, my stack, my rules about not touching audio delivery. Nobody sells that. Nobody can, it's mine.

What it cost me was about 5 hours of hands on time, and a weekend's worth of tokens on my normal $20 a month Claude account. The same account I already pay for, for all the other things I build. Not a new SaaS bill. Not a new vendor relationship. Not a contract I'll forget I'm paying for in fourteen months.

I want to be honest about scope here, because I don't want to oversell it. This was not a complex system. It was a focused, well-defined tool solving one problem for one server I control. All told it's about 5,000 lines of bare metal PHP code plus the unit tests. No frameworks, no Composer packages.

I'm not telling you I refactored a payments platform in a weekend. I'm telling you I built a small, real, production thing, tested and live, in the time it usually takes me to watch the University of Tennessee football team beat the stuffing out of a community college team and the Dolphins lose to somebody I don't care about.

And that's exactly why it matters. Because "small, well-defined tool that would normally cost $10 to $20 a month forever" used to mean one of two things. Either you paid the SaaS, or you carved out a real chunk of a real month or more to build it yourself, tested it, and hoped you got it right. Neither of those was a weekend and a coffee.

The build versus buy math just moved. Not for everything. Not for the complex stuff, not yet anyway. But for the small, well-scoped, "I know exactly what I need and nobody sells exactly that" category of software, the calculus is different than it was two years ago. Worth sitting with that for a minute before you reach for the credit card on the next $15 a month tool.

Okay. That's the story. Now go build something small and see what it costs you.




For the record, the two bad movies were:

Kathy was right to skip both.