TubeRefine

AI

Find what's worth watching — and know it before you press play.

A few taps via YouTube's share menu triggers automatic LLM tagging, while the "Generate Summary" button delivers detailed insights on demand.
Effortlessly manage content by genre, channel, and length, then keep your library clean using 5-star ratings and soft deletes.
Powered by an integrated translation engine, multi-language search lets you query global videos in English, regardless of source language.

TBD

Try demo GitHub ↗

01Problem → Solution

PROBLEMSOLUTION

PROBLEM 01

The gems get buried in the noise

Algorithmic suggestions are anchored to your watch history, lining up more of the same. What you end up watching gets decided with no regard for your actual intent.

SOLUTION

Break the dependence on the algorithm and hand control back to a library of your own, kept outside YouTube.

PROBLEM 02

Watch Later in name only

The native Watch Later has no search and no filters. The more you save, the deeper videos sink, locking you into a pile you hoard but never watch.

SOLUTION

The YouTube API plus AI-generated tags make every saved video searchable, filterable in an instant by topic, length, recency, and more.

PROBLEM 03

You can't tell if a video is worth it until you open it

As mass-produced, low-quality videos pile up, the only way to judge one is to open it. Opening, checking, and bailing out each time drags down how efficiently you can watch.

SOLUTION

Call up an AI summary on demand and settle the watch-or-skip decision without ever opening the video.

Bulk-generating summaries for every saved video was deliberately left out. The core was fixed on two things: being able to search and retrieve, and being able to judge without opening.

02Key features

01

Multi-axis filtering and high-precision library search

Auto-taggingAI generates supplementary tags in the video's own language, then auto-translates them to English for search — on top of the YouTube API data
Multi-condition filterNarrow down by combining topic, length, publish date, and rating
Cross-language searchSearch titles, descriptions, and tags at once, in any language

DESIGN RATIONALE

AI-generated tags put searchability itself at the core.

02

One-tap AI summaries and pre-watch judgment

One-tap summaryGenerates summaries on demand with Gemini 2.5 Flash
Per-chapter timestampsChaptered videos display with links to each relevant point
Judge without openingSettle whether to watch from text alone

DESIGN RATIONALE

Judging a video without opening it is designed here as a new kind of viewing experience. Even for thin, mass-produced videos, it offers the experience of "reading" a video as text.

03

Seamless saving from the YouTube app

Share-menu integrationLaunch the save screen straight from YouTube's share menu
Instant tagged saveAdd custom tags and send it straight to your library

DESIGN RATIONALE

When there's friction between discovering a video and saving it, you end up thinking "I'll add it later" and your library never grows. By putting the starting point for saving inside the YouTube app itself, the design keeps the library fresh and complete.

04

Organize videos you've finished, rating as you go

Soft deleteDeletions move to a trash bin to prevent accidental loss
5-star ratingBuild up your post-watch ratings over time
Multi-axis sortingOrganize your library by rating, date, and more

DESIGN RATIONALE

The longer you use a library, the more it bloats, mixing in low-quality videos. By designing deletion, rating, and sorting as one set, it defines a way to manage a library you can keep refining even as it grows.

03Tech stack

TBD

04Why I built this

Using YouTube as a podcast platform revealed a key limitation:

Saving videos is easy. Finding and vetting them isn't.

YouTube's default Watch Later offers no search or filtering, turning saved videos into a black hole. Worse, checking relevance required scrubbing through videos—a high-friction time sink.

I built this app to solve both problems. By combining YouTube API metadata normalization with an on-demand Gemini pipeline, it turns saved videos into a searchable, text-first personal library.

Instant Access: Quickly locate and retrieve saved videos when needed
Text-First Consumption: Read AI summaries to bypass video playback entirely
Intentional Curation: Choose what to process without algorithmic bias

— The Outcome —

Shifted from passive watching to intentional, text-first learning—skipping videos whenever text suffices.

05Engineering decisions

01

Chose Gemini 2.5 Flash as the AI engine, decided by its unique ability to access videos directly from a URL

PROBLEMOther LLMs need to fetch and process captions separately, and can't handle videos with no captions

Other models, even when handed a YouTube URL, can't reach the video itself, let alone its audio or footage. Gemini was the only model that accesses audio, video, and captions directly from a URL. Weighing the balance of cost, response speed, and accuracy, I settled on Gemini 2.5 Flash.

Any video can be summarized, regardless of whether it has captions

02

Made summary generation on-demand rather than automatic on save, eliminating wasted API cost

PROBLEMNot every saved video actually needs a summary

I considered auto-generating on save, but decided not every video needs a summary. Instead, it generates in one tap at the moment the user actually wants it.

Cuts unnecessary API calls to zero while still letting you read instantly the moment curiosity strikes

03

Structured tags in three layers — API, AI, and custom — putting searchability at the product's core

PROBLEMMany videos have no YouTube API tags at all, leaving search coverage full of gaps

Since API tags alone can only cover so much, I layered on AI-generated supplementary tags and the user's own custom tags for a three-tier structure. Tags are unified in English, and at search time the user's input is translated to English in the background before matching.

A library where you can pull up any video without a miss, without thinking about language

Try it out

No signup required — guest mode available

Try demo GitHub ↗

← LinguaCoach ↑ Back to Work