Welcome to the new rubric of my newsletter where I share the weird stuff I make! It’s remarkable how we’ve survived 15 letters without it given my affinity for side-projects.

Problem to solve

From time to time, I get carried away by a topic (airships, remember?). I want to know EVERYTHING about it. This lasts from a couple of hours to a couple of decades.

Yet when I try to find videos and podcasts on the topic (I listen more than I read), the top results are usually pretty bad — especially when there’s some fucking SEO commerciality involved:

At the same time, I’ve accumulated a rather long list of creators I trust (here are some). I wish I could search their content first before rummaging through the entire sad internet.

So here’s what I have in mind. I tell a bot or an agent what I’m interested in, and it suggests content from a curated list of creators.

Solution A: Claude skill (~1h)

Type gcc <topic> and Claude does the job:

Once I understood the ecosystem, this was quite easy to achieve. Here’s my SKILL.MD.

However, the experience is meh. Claude doesn’t show thumbnails and hates it when I try to open links (the only purpose of this whole thing):

Claude is relatively slow and unpredictable:

As well as simply unstable:

So I’ve decided to build an old-school deterministic Telegram bot using the cutting-edge vibe coding agentic engineering technology.

The guys on Twitter say that coding is solved and claim to have built an entire browser without a human in the loop. Certainly, my little bot is a 20-minute adventure. In and out.

Solution B: Telegram bot (~24h 🥲)

Three full days later, here we go:

Submit a topic to the bot — and it semantically (!) searches all your YouTube and podcast subscriptions in seconds.

Here’s the bot’s source code. The whole thing can be hosted on a small server with 2GB RAM, here’s my referral DigitalOcean link.

If you want to try the bot but couldn’t be bothered hosting it, let me know. Perhaps, I can make it multi-player. Maybe. No promises 🙃.

Is the future here?

Well, yes and no.

✨ Unlike with Flag Quiz where I had to manually rewrite sections of code because of undebuggable race conditions (May 2025), I haven’t written a single line of code for this.

✨ I dictated, not typed, almost all my prompts with Superwhisper. This removed friction and encouraged thorough explanations. The speech to text is (finally) robust enough, and Claude doesn’t care about the rare mistakes.

💩 It still took a lot of time to get things working smoothly. Getting to 80% was a walk in the park, getting to 100% was an Apollo mission. I spent 4 hours fixing memory consumption alone, not a fun experience.

💩 For the love of god (Amo·dei), I couldn’t make Claude autonomous enough. Even when I tried to build loops for the agent to self-correct, it failed to do so most of the time. The whole project required a lot of human direction.

A year ago, this project would’ve taken me a week. Now it took three days (and I optimized for learning, not speed). An improvement? Fo shur. The bar for tinkerers like me is lower now, and it’s enticing to try projects that were deemed unfeasible before.

At the same time, “the era of bespoke software” is not here. Not yet.

What would make me reconsider? Falling in love with a product and then finding out it was made by a single gal and her team of agents in a week.

Until then, I’ll be rolling my eyes reading industry Twitter, sorry.

This was issue #16, find more at antoniokov.com.

What’s been your experience building with AI? Reply to this email and share your thoughts (and questions?).