I consume most of my content through video — YouTube, Instagram, random lectures I find at 1am. But the tools I actually think with, like ChatGPT and Claude, don't watch video. So there's this constant gap: I watch something that genuinely changes how I think, and then there's no easy way to actually sit with an AI and talk it through.
That annoyance is where this started. Let me just build something that turns video into text, so I can chat about it properly.
The first version — the obvious one
Paste a YouTube or Instagram link, get a transcript, chat about it. Pull captions where they already exist, fall back to Whisper transcription when they don't — especially useful for noisy audio or platforms that don't offer captions at all.
Then I started thinking through how I'd actually use this day to day, and the product architecture started to change with it:
- Embed the video right inside the app, so I never have to jump somewhere else just to get a transcript or start a conversation about it.
- No algorithmic feed. I search for and follow specific channels — no infinite scroll, no thumbnail bait pulling me into a two-hour rabbit hole I didn't mean to enter.
- For discovering new creators, still no direct recommendations — just similar creators and their most-watched video on a topic, sitting in a tab I visit on purpose, not something pushed at me.
I trust an AI answering my actual question more than an algorithm optimized to keep me watching.
YouTube's recommendation engine doesn't care if a video is good for me — only that I keep scrolling. An LLM I ask something directly isn't rewarded for keeping me around longer. That difference is the whole point.
The reality check
Then I actually looked around, and it stung a little. "Video to transcript, chat with the video" already exists, and it's crowded. Tools like Glasp have millions of users doing exactly this, across multiple AI models, for free. Competing on better summaries means competing on something already commoditized, against players with a massive head start. That's not a moat. That's just table stakes.
So I had to sit with the harder question: if the core feature isn't special, what am I even building?
The realization that changed everything
Here's the pattern I kept running into, and it's uncomfortably relatable: knowing something is bad for you doesn't change your behavior. People know YouTube is designed to hook them, the same way people know a lot of their habits aren't great for them — and knowing changes almost nothing. Awareness was never the missing piece.
What might actually work is replacing the loop, not just deleting it. And that's where the real idea showed up:
A private, reflective diary for what I watch and think — not just a pile of transcripts, but something that holds me accountable.
Say I watch five videos on phone addiction over a few weeks. Instead of just storing five transcripts, the app checks in with me: "You've watched five videos on this topic, covering roughly X, Y, and Z. Did you actually act on any of it? What worked? What got in the way?"
That's the piece I haven't seen anywhere else. It closes the loop between consuming something and actually doing something about it. It also creates a layer most summarizers don't capture — not just what the video said, but what I thought, asked, challenged, and came back to over time. A record of how my own thinking has moved, built from interaction rather than passive consumption.
That's the moat, if there is one. Not sharper summaries — a private, compounding record of my own thinking and follow-through, private enough that I'd actually be honest in it. The second it becomes visible to anyone else, even anonymously, people stop reflecting and start performing for an audience — which is exactly why "see what others went through" has to stay a separate idea for much later, not something baked into the core.
How I'm planning to get there
Test it on myself, no code involved
Pick one topic I actually care about, watch 5–10 videos on it over time, talk it through with an LLM after each one, then try writing my own "here's how my thinking changed" summary at the end.
Sharpen the implementation layer
Turn the CS fundamentals into product-building fluency: Python, APIs, the terminal, data flow, and the tooling needed to move from a technical idea to a working system.
Build the smallest working pipeline
Design the first end-to-end path: ingest a link, resolve available captions, fall back to transcription when needed, and persist the resulting content in a form the product can actually use.
Turn the system into a product
Wrap the pipeline in a focused interface: paste a link, retrieve the source, surface the transcript, and make the conversation happen alongside the original video.
Build the actual wedge
Group things by topic. Build the check-in — a proactive nudge that sums up what I've watched and asks what I actually did with it. This is the hard part, and also the whole point.
Only if stage 4 actually works
Explore a discovery tab with no algorithmic feed, and maybe, much later, a separate opt-in space for anonymized peer insight — kept deliberately apart from the private diary.
I'm deliberately keeping the scope tight: each layer has to earn the next one. The goal isn't to ship a lot of code; it's to find out whether the product creates a habit worth building around.