NexGen News · 2026 · Solo · 10 hours · Exploration · Not yet shipped
News makes you start over on every visit: one story, dozens of headlines. In about ten hours, solo, I built the research engine first, then designed around one question instead of a feed: news that remembers what you're trying to understand and shows only what changed. A working prototype.
So I built the intelligence before the interface, and spent ten hours on why it should exist.
And every visit, you start over.
Don't give people more news. Help them uncover the story.
Watch the agent work. Tap a cluster to see what it's reading.
Real sources, every quote checked, memory kept in an Obsidian vault.
No spinner. Explore what others already uncovered.
Each file is one development, with its sources, its dates, and how far the evidence goes.
No feed. One question, and what changed since you left.
Mark what you already knew. It remembers, and never repeats it.
Next: live inquiries, tested with real people. I'd measure one thing: connections they say they would have missed.
For a year I searched for a better AI-native workflow. Then a friend introduced me to one person's method, and it changed my craft. The answer already existed. I had no way to find it.
You don't know what you don't know.
One question. Less noise. More understanding.
News moves on and you start over. Every visit, you rebuild what you knew and guess what's genuinely new.
One story in five framings overloads attention, and the conflicting takes create dissonance. People resolve it by tuning out, or by keeping the version that already agrees with them.
37% trust news most of the time. 20% trust news from AI chatbots. The brief asks AI to repair a trust problem. People trust AI news even less. (Reuters Institute, Digital News Report 2026)
Don't give people more news. Help them uncover the story. Say what you're trying to understand. NexGen News cuts through the noise and connects millions of dots into a meta perspective made for you.
Saying a question is faster than typing it, so voice comes first. Typing is the fallback. Every update answers: what's new, why it matters to you, what you're missing, and the evidence.
Sources gather as a living field of dots while your inquiry progresses. Tap a cluster to see what the agent is looking at as it scans the web.
Each inquiry fetches real public sources, checks every quote against the original text, and writes what it learns into an Obsidian vault. The next visit starts from that memory.
No guesswork. Designing on top of a working engine grounds every UX and product decision in what the system can actually do. The graph is the real vault after four inquiries: 180 notes, 243 links.
The inquiry is real and takes several minutes. Instead of a spinner, you browse closely related files other people have already compiled. It's also a good moment to show, step by step, how the app works.
Community files show how others connected the dots around closely related questions, and may already deepen your understanding. Open your first inquiry when you're done exploring. Community files always wait at the end of your own.
Each file is one development, with its sources, its dates, and how far the evidence goes.
No feed. Your question, your last update, and one button to check what changed.
I reached the 10-hour limit here. I focused on onboarding and a preview of the next visit, so the screens from this point are unpolished. I'd replace the "Check what changed" button with a different interaction pattern.
Each story says how far its evidence goes. Mark something you already knew and it won't come back as news.
The vision: the engine searches for new reporting on your inquiry once a day, in the background, so something new arrives without effort. A native notification says when it's ready.
Your marks shape the next update, so nothing repeats. Each visit brings a fresh perspective on what you want to understand.
Next, turn on the inquiry engine so searches run in real time. Then test it with real people. Or keep iterating until every tester loves it and asks for more.
How I'd measure it. The test that counts: does it surface a consequential connection you would have missed, with evidence you can inspect? Then run two bets side by side, a reliably fresh briefing or a surprising perspective, to see which people value. Measure feeling caught up and return visits, never minutes in the app.
Solo: working research engine, product, interface, motion.
Not shipped. No user interviews yet. The returning-user experience is unpolished. X & YouTube block automated access.
I hit the 10-hour limit at Home. I traded the return visit for a first impression, on purpose.
Next case: The agent lives in your inbox