Riffle Feed
Watching someone demo a tool in a video is rarely how people retain workflows. Most of us learn when we sit down to solve a specific problem in our own work. Riffle turns hundreds of hours of recorded tutorials into clean, searchable reference cards on your phone—grounded in source transcripts and linked to the original videos.
Small LLM Dispatchers: Fast Routing Before Heavy Reasoning
- Avoid sending unvalidated user inputs directly to expensive 70B+ reasoning models.
- Use an 8B local model to classify intent and validate in <120ms.
- Route queries to the heavy model only when multi-step synthesis is genuinely required.
SOURCE-GROUNDED
Checking links against actual downloads
When an AI model summarizes video transcripts into cards, it frequently guesses YouTube links and sends readers to the wrong video. We cross-reference every card against our raw download logs. If a link cannot be verified directly, the card gets flagged immediately.
LEARNER CONTROL
Tapping words you don't know
When we tested letting an AI automatically build a glossary of unfamiliar terms, it defined basic words everyone understands while skipping technical jargon readers actually needed. With tap-to-flag, you tap any word you don't recognize directly on the card to save it to your own study queue.
DESKTOP HAND-OFF
Compiling cards into ready-to-use prompts
Reading summaries on a phone is passive. When you star cards in Riffle, you can compile them into structured research prompts tailored to your specific work. When you get to your computer, you copy the prompt directly into ChatGPT or Claude to do the actual job.
Why This Exists
Watching someone else build something is mostly television.
Watching someone build a workflow in a 20-minute video feels productive while you watch it. But because you don't need that specific tool that day, you rarely retain the steps. Three months later, when you actually face that problem in your own business, you can't remember which video had the answer or where the creator showed the setup.
Generic AI summaries don't solve this. They smooth out technical nuances, hide edge cases, and leave you with bullet points that don't help you build. And when you sit down at your computer three weeks later, you still don't know what to type into a prompt.
Riffle was built to solve that retrieval gap. It ingests video transcripts, cleans the metadata, organizes the material into clear topics, and compiles concrete steps into structured reference cards on your phone. Instead of endless scrolling, you save what looks relevant, organize cards into custom decks, and compile them into targeted prompts when you sit down at your desk.
Current App Snapshot
What's active in the app today.
Want to turn your own company recordings, calls, or training into a private, grounded knowledge base in Google NotebookLM?
Learn about the Knowledge Pipeline SystemUnder the Hood
Six practical mechanisms that keep the feed reliable.
Riffle is not an automated scraper or an open-ended chatbot. It is a structured reading tool built with strict boundaries.
Download Log Matching
Cross-references YouTube video IDs, start offsets, and channel handles against raw download logs to prevent broken or invented links before cards reach the feed.
5-Mode Prompt Compiler
Turns starred cards into structured prompts across five thinking modes—from business application to deep research—ready to paste into desktop AI tools.
Tap-to-Flag Lexical Capture
Lets readers tap words or phrases directly in card text to queue unfamiliar jargon for personal study, replacing unreliable AI glossary guesses.
Lightweight Feed Rendering
Renders only the active card and immediate neighbors rather than mounting 500+ cards at once, keeping mobile scrolling fast and preventing browser crashes.
Your Data Stays on Your Device
Stores reading history, starred cards, and flagged terms locally in your browser. Includes clean JSON backup export and import with zero tracking.
Non-Mutating Review Staging
Stages mobile review exports into separate curation queues for deep research, term definitions, and card updates without directly overwriting master catalog files.
Challenging Our Own Claims
What we hoped for, what we tested, and what actually happened.
It is easy to assume an AI feed will work nicely until you use it on a phone every day. Here are four assumptions we had to throw away during development, and the real solutions we engineered to fix them.
When the AI guessed YouTube links and got them completely wrong.
Why we had to build an automated link checker to stop the AI from inventing fake sources.
We assumed that because Google NotebookLM has access to the video transcripts, its citations would reliably link back to the correct original videos.
It didn't. In raw test batches, the AI frequently guessed video links. It would mix up video codes, mangle web addresses, or attach the exact same YouTube link to three completely different tutorials from different creators. If someone clicked a card to verify a step, they were sent to the wrong video.
We refused to trust the AI's link guesses. We built an automated checking script that intercepts every card, looks up the video title in our original download logs, and replaces the AI's guess with the verified video link. If a link cannot be verified directly, the system flags it immediately. A wrong link is worse than no link.
Why reading summaries doesn't help you build (The Prompt Compiler).
Turning passive card reading into ready-to-use prompts for ChatGPT, Claude, and NotebookLM.
We initially thought that giving people short, concise summaries of 20-minute tutorials would be enough to help them learn new tools.
It wasn't enough. People would read a card on their phone, nod along, and then sit down at their desk three weeks later still not knowing what to type. A summary tells you what someone else did; it doesn't help you do the work yourself.
We built a prompt generator directly into the app. When you star cards on your phone, Riffle can compile them into structured research prompts tailored to your specific work. You choose whether you want the AI to "Apply this to my business," "Conduct deep research," or "Explain like I'm new." When you get to your computer, you copy the prompt directly into ChatGPT or Claude to do the actual job.
Why asking an AI to guess "unfamiliar words" never works.
Letting the reader highlight unknown terms instead of trusting an AI glossary.
We thought we could use an AI model to read through technical video transcripts and automatically highlight all the "confusing terms" for the reader.
The AI did a poor job guessing what a human reader actually knows. In pilot testing, recall topped out around 45%. The model spent time defining basic terms everyone understands, while highlighting self-evident card titles and skipping over the specific technical acronyms that readers genuinely needed explained.
We put the reader in control with a tap-to-flag feature. While reading any card on your phone, you can tap and highlight any word you don't recognize. Riffle automatically saves it to your personal learning queue so you can review it later or include it in your next research prompt.
What happened when we put 500 cards into a mobile browser at once.
Why mobile browsers freeze on long feeds, and how selective rendering fixed it.
We assumed modern mobile phones had plenty of memory to handle a list of 500 text cards with smooth snap scrolling.
The mobile browser struggled immediately. Rendering hundreds of cards—complete with formatting, interactive buttons, popovers, and gesture listeners—exhausted mobile memory, dropped scroll rates, and caused the interface to lag or freeze entirely.
We separated the scroll frame from the heavy content. The app keeps a lightweight scroll container, but only renders the card currently on screen plus the cards immediately before and after it. Off-screen cards unmount until you scroll near them, keeping phone memory low and scrolling responsive.
Working Demonstration
Turn your own company recordings into a searchable library.
Riffle is a public demonstration of what happens when you structure video recordings properly. If your team has hours of recorded client calls, training sessions, or internal walkthroughs sitting untouched in folders, the Knowledge Pipeline turns them into a private, searchable reference system inside your own Google accounts.