You've no doubt had some wins by now with AI. Drafted something faster. Extracted key data from some dense PDFs. Cleared a bottleneck that would have eaten your afternoon. You did it and it made sense, right?

You felt in control because you understood the assignment. And it wasn’t prompting tricks or rigid structures. It was just clear communication of intent.

And once that clicks, something shifts. Like a little radar popping up in your peripheral vision, you start measuring every task that crosses your desk as a candidate for AI enrichment. That's real. That's progress.

But what about the problems beyond text or file organization? The daily frictions that don't need a prompt — they need software. The "if only there were an app for that" moments we all have.

Here's the question: Do you believe you could actually build your own app to solve your own problem?

At this point, it really shouldn't seem alien. Because the same skills you're already flexing by thinking clearly about what you want, describing it, and iterating through conversation, are the exact same skills you need to build software in 2026.

You don't need to code. That IS the revolution. You just need to understand the problem and have the judgment to recognize when you've solved it.

Let me show you some things I built over the last few months to solve real pain-points I hit while building this newsletter and my consulting business. Each one forced me to answer the same questions differently: What is the problem? What would solve it? And what does "done" look like?

Those questions are the ballgame, folks.

FastAPI/Python/PWA

We remember it exists. Might even remember the channel. But by the time you scroll through a long watch history and scrub through 45-minutes of video hoping to recognize the sentence when we hear it, you’ve basically diverted into a whole other project.

The knowledge was useful, but fleeting.

YouTube provides transcripts, but it’s almost impossible to copy them from your phone. There are other apps that create YouTube transcripts too.

But I wondered if I could build something that went further. Not stopping at just the transcript. After all, having that only solved half my problem.

I wanted an app that captures YouTube videos I flag, transcribes them automatically, and parks that text somewhere I can actually find it and search it. Better yet, somewhere an AI agent knows it’s there and can synthesize it into my knowledge base.

I brought the problem to Claude. To get started I gave it barely more than the preceding paragraphs you just read, and then we discussed my options.

The first attempt was about keeping it as simple as possible. I was aware of Apple’s Shortcut feature that makes it easy to automate multiple actions in steps. The idea was, I copy a YouTube URL into the shortcut, its automation writes that URL to a text file in a shared iCloud folder.

We tested it. It worked.

So then we built a small Python script that ‘watches’ that shared folder from my Mac, polling it every 5 seconds looking for a new entry. When it detects one, it uses a second script to extract the URL and pass it to a third script that fires up a little software nugget called yt-dlp (an open‑source tool that downloads from YouTube and other video platforms). To keep our file sizes small, we only download the audio stream from YouTube.

Once the audio file is downloaded it gets passed to FFmpeg, a free, open‑source tool that is like a Swiss army knife for media file type conversions. It renders it into an even smaller .m4a file.

The next script delivers that audio file to Whisper — OpenAI's free, downloadable speech-to-text model, running on my own machine — nothing sent to the cloud.

Whisper transcribes the audio to a text file and that file is detected by the next script, which deposits it to a specific database in my Notion workspace where my AI agents can read it, parse it, extract themes and ideas, and synthesize it all into my knowledge base.

Well guess what? It worked. Right up until it didn’t.

The architecture was sound, but YouTube has this little dance they do, periodically rotating some wiring under the hood that breaks the yt-dlp piece. It’s an easy fix. Takes a few seconds really. But knowing when you need it is unpredictable. I didn’t realize until I checked my Notion database a couple weeks later and a bunch of videos I thought I transcribed weren’t there. Not cool. But a clear problem.

I needed feedback. Visual feedback, telling me if a transcription was a success or not.

Apple shortcuts is not up to that task, so Claude suggested a different backend framework and we vibed up VID2NOTION v2. The difference was mainly an interface upgrade. Now I could paste the URL, hit send, and watch as the status updated from queued to processing to done. If something fails, it says so. I can retry with one tap.

Express/Node.js/Gemini Vision

Flea markets. Estate sales. The kind of places where you're holding something old, interesting-looking, and you have about thirty seconds to decide if it's a $5 novelty or a $200 find before somebody else picks it up. You can try to Google it, but what do you type? "Old green glass bowl with flowers"? By the time you've scrolled through results, the moment is gone.

This idea was born over breakfast with a friend, waxing nostalgic about our shared love of flea markets. It kicked around my fun-projects list for a while. Then I thought: what if I could just point my phone at something and quickly discover its actual value? Not a Google search — an actual identification. What is this thing? What's it worth?

The build: Gemini 2.0 Flash for AI Vision — take a photo, the AI identifies the object and suggests what to search for, and not just "that's a bowl." We wrote the prompt to push Gemini to identify items the way an experienced picker would — exact pattern name, era, size, material, distinguishing features. That’s called a schema.

A piece of Pyrex doesn't come back as "vintage glass dish." It comes back as "1970s Pyrex Butterfly Gold casserole dish with lid, 2.5 quart." That specificity is what makes the next step work.

The detailed identification is then fed to a service called SerpAPI that can search eBay's completed sales data. Not current asking prices — actual sold prices, with links to the listings. The app calculates low, average, and high prices from real transactions, not algorithmic guesses.

About 200 lines of code. Photo goes in, identification comes back, pricing follows. We had a working prototype by mid-morning, day one.

That's where the story should end.

It doesn't.

This app was mostly for fun. There was no deadline. No urgent need. So as I casually imagined myself roaming the stalls of a flea market using it, my addled Gen X imagination quickly morphed that into a scene akin to Dr. Who rummaging through Diagon Alley if it were on Mos Eisley.

So naturally I followed this muse and decided the app needed a full sci-fi steampunk user interface, like a physical machine from a time out of time. Old oscilloscope vibes. Green phosphor text on dark backgrounds. Gauge displays for the price range. A cylinder-style ticker that scrolls through recent sales. Status lights that illuminate as each step completes.

The lesson: Know what "done" means before you start polishing.

Again, I wrote no code myself. The conversation built it. The fact I went rogue on a steampunk fantasy UI side-quest and derailed the project was all on me.

I think I’m funnier than her. She’s got a great poker face, but I think I make her laugh on the inside.

Here’s what our “office” looks like. Crazy, right? But it works and I love it.

Want one?

Thanks for reading,