Shelf

Tracking five kinds of media across five different apps — and none of them recommends across formats or reliably follows releases.Tracking five kinds of media across five different apps — and none of them recommends across formats or reliably follows releases.

Manga in one app, films in another, books in a third: progress scattered, and nothing bridges formats or tracks manhwa releases well. AniList doesn't know my Letterboxd, Goodreads doesn't know my manga, and none of them recommends across formats. I wanted one private place for everything I read and watch, with an assistant that actually knows my library.

A personal, single-user app — but held to product standards: real catalogs, honest release data, and latency low enough to feel instant.A personal, single-user app — but held to product standards: real catalogs, honest release data, and latency low enough to feel instant.

• Real data, not a toy. Recommendations and tracking pull from live catalogs — TMDB, AniList, Google Books — so the assistant has to search sources I don't own yet, not just read back my own shelf.
• No perfect release API. No free source cleanly covers manhwa release dates, so the tracker has to stay right even when any single source is wrong or down.
• Latency. Recaps and chat only feel usable if they stream and stay fast, which drives the model choices rather than the other way around.

A Claude companion that tool-calls real catalogs, a resilient multi-source release cascade, and a streaming layer that picks the right model for the job.A Claude companion that tool-calls real catalogs, a resilient multi-source release cascade, and a streaming layer that picks the right model for the job.

The core is an AI companion: a streaming chat (NDJSON) that calls tools to search real titles, understands a persistent taste profile, recommends toward or away from it, filters by streaming platform, bridges anime ↔ manga ↔ book, and returns actionable cards (add / where to watch). Release tracking runs a ComicK → MangaDex → MangaUpdates cascade with guardrails — never showing a number below what you've already read — plus cadence estimation, so chapters stay accurate with no manual upkeep. Recaps and chat stream, with Claude Haiku by default for speed and Sonnet on demand; thinking is disabled to cut latency, with a fallback on overload. Underneath: Next.js 14 App Router with RSC (force-dynamic) and Server Actions, a Prisma 7 / PostgreSQL core, Supabase Auth on an allowlist, and a bespoke "liquid glass" design system with a command palette and anti-CLS skeletons.

Around sixteen sections orbiting one assistant: library, discover, calendar, a seasonal-anime hub, a "what do I watch tonight" decider, and a two-person mode.Around sixteen sections orbiting one assistant: library, discover, calendar, a seasonal-anime hub, a "what do I watch tonight" decider, and a two-person mode.

A working private tracker — home, library, discover, calendar, stats, chat, a shared "à deux" mode, a seasonal-anime hub, and detail pages: roughly sixteen sections, all served by the companion. Spoiler-free recaps, a "tonight I'll watch what" decider, ISBN barcode scanning, a release calendar with Google Calendar export, and push notifications round it out. It's a single-user app I actually use daily, not a demo.

The multi-source cascade earns its keep — but the cross-source matching layer is where the next reliability work lives.The multi-source cascade earns its keep — but the cross-source matching layer is where the next reliability work lives.

The provider cascade was the right call: resilience beat betting on any single "best" source. But the hardest bug class was wrong chapter numbers from bad title matches across sources, patched with guardrails rather than solved at the root. I'd invest next in a stronger cross-source identity layer, so a title resolves to one canonical entity before any number is trusted. I'd also lean harder on the taste store: grounding recommendations in it and excluding owned or dismissed titles fixed the "recommends what I already have" problem, and there's more signal in there to use.