I've been reading about the Iran conflict every day, refreshing the same sites. Trying to imagine where its going. And with all the vibe-coding going on, I built a thing that does it for me.
It's at Pulse. It's a (not very good) personal intelligence briefing that uses Claude to search the web, synthesize what's happening across six domains (military, humanitarian, energy, diplomacy, domestic politics, second-order effects), track market indicators, and... make predictions. That last part is the fun part.
I it's just a test, a proof of concept? A toy? My own custom geopolitical analyst? It's all of those, maybe. I built it in a couple days and I've been running it 3-4 times a day.. It gives me a structured briefing every time, tells me what's actually new, and tracks its own predictions over time. Right now it claims it's running at 75% accuracy on resolved predictions. It correctly called Brent crude going over $125 before March 14 although I think that was just futures markets temporarily. It has 8 more predictions pending. We'll see.
There's something weirdly compelling for me about watching an AI try to guess what happens next, then checking if it was right. I was thinking about adding GPT5.4 or whatever to it so I can watch two models disagree with each other and see who's better at forecasting. Haven't done that yet but I want to.
The first time I ran it, the AI agents went out searching the web and racked up over a million input tokens in one go. Which sounds like a lot because it is a lot. Anthropic's rate limits are 30,000 tokens per minute, so it immediately hit the wall and I had to wait. I ended up spreading the search calls out and using Haiku (the smaller, cheaper model) for the web research phase, and then Opus for the actual analysis.
That worked, mostly. But Haiku made stuff up sometimes. It pulled gold prices from like 2024 and presented them as current. It fabricated some numbers that threw off the market analysis. I had to add sanity checks into the prompt ("gold is above $5,000 in 2026, if you're seeing $2,900 that's wrong"). Fortunately Opus does the final synthesis so the actual briefing is solid, but it was a good reminder that even scarping with cheaper models cut corners in ways that aren't always obvious.
If you think about the implications of building a real media business this way... it would be expensive. Every scan costs real money in API calls, and you need the good model for the parts that matter. You can't just Haiku your way through everything.
A couple friends I shared it with keep checking it now, which I didn't expect. So that's interesting to me. The idea that you could build a personalized news product for a small group of people who care about a specific thing. who knows.
Final thing I might add a fake newspaper version of it because the interface is a bit clunky. Same Claude-sourced analysis, but laid out like a broadsheet with an opinionated columnist voice. An editorialized AI newspaper. That sounds fun and also possibly unhinged. I might do it.
For now Pulse is just a thing I built for myself that turned out to be more interesting than I expected. If you're following the Iran situation, take a look. The predictions section at the bottom is worth scrolling to. And if you have thoughts on it text me.