Content Hub

What this POC proves, and what it surfaced

• Ingestion works with no API key, across three feeds. All 45 items came from free per-channel RSS: Fox International UC7l_C-SB-TyxXR3qmgBnWPw, Korda TV UContEaeqonv1e131k03FyMA, Fox Rage UCGetpCv4RI4x5YEnK3luu1w. Each returns id, title, description, thumbnail, published date, view count and rating. One parser, three suppliers, no key, no quota.
• Playback is the standard embed. Click any thumbnail: it swaps to a youtube-nocookie.com iframe. Nothing is rehosted, the view still counts for the supplier, and we stay the right side of copyright.
• Every supplier has a completely different content profile. This is the big finding. Three feeds, three shapes, and a single global rule would fail all of them:
  • Korda is a broadcaster. 14 of 15 long-form films, one Short. Views 55k to 378k. Named series (Thinking Tackle, Mindset, Underwater). Almost no product launches.
  • Fox International is mixed. 5 long-form, 10 Shorts. Views 2k to 99k. Films plus two big product launches.
  • Fox Rage is a shop window. 14 of 15 are Shorts, one long-form launch. Views 547 to 5,329, roughly 100x below Korda. Nearly every item is a product ad.
  Consequence: per-source configuration is not a nicety, it is the design. Review effort, cadence and default treatment should all be set per feed.
• "Promo" does not mean "worthless", and that breaks my earlier proposal. I previously suggested auto-suggesting promo for rejection. Fox Rage kills that: its promo Shorts are its product content, each naming a specific buyable item (Prism X Power Spin, Voyager XL Mat, Belt Pliers, Scent Shad). Auto-rejecting promo would cut Fox Rage from 15 items to 1, and it is the most directly shoppable feed of the three. Promo needs splitting into product-promo (keep, deep-link it) and filler-promo (drop, e.g. "June 16th river madness!", "Kiss the camera Moz!").
• Shorts still must not be filtered by format. Unchanged and now stronger. Fox's "60 Seconds to tie a Solid Bag" is a real tutorial on 86,860 views in Short format, and all of Fox Rage is Shorts. A format filter would bin both. Toggle "Hide promo" to see it: 45 drops to 21, Korda barely moves (15 to 14), Fox halves (15 to 6), and Fox Rage collapses (15 to 1).
• Discipline is a channel property, but the override is real. Mostly confirmed: Fox International is carp, Korda is carp, Fox Rage is predator. But Fox Rage posts coastal wrasse and creature-bait content to its predator feed. Those two items are filed Sea here. So default_discipline per feed is right, and a per-item override is required, not optional.
• Series is a missing field. Korda organises around named franchises (Thinking Tackle, Mindset, Underwater, Kontender, Spooners Vlog), Fox around The Weekend Mission and Carp Fishing Edges. Anglers follow series, not just brands. series should be a first-class column, and a likely "follow" target in the logged-in slice.
• Regional duplicate channels are a trap. Fox Rage also runs Benelux, España and Sweden channels carrying the same launches in other languages. The pilot must pin the UK/main feed per brand or we ingest each product launch four times.
• The brand logo library already exists, with one gap. TC hosts brand logos on the Shopify CDN under a clean convention, <Brand>_Tackle-City_brand-logo_1350x1350.png, currently covering FOX, Korda, Abu-Garcia, Berkley, DAM, JRC, MADCAT, PENN, Savage-Gear and Ugly-Stik. So card logos resolve by name with no new asset work. Fox Rage has no logo in that library and borrows the FOX mark here. Worth adding, since Fox Rage is a distinct brand to a predator angler.

Ask, Kit and Wish Lists

• The Ask box is a stand-in, and honest about it. A static page cannot call Haiku, because an API key must never ship to the browser. The concept dictionary in the source approximates what the model would return. In production the query goes to a Worker, Haiku turns the question into structured intent, and retrieval runs over pgvector. Showing "Understood as" is deliberate: the interpretation layer stays visible and checkable rather than magic.
• Reuse the embedding model already pinned. BUILDSPEC-03b pins OpenAI text-embedding-3-small (1536 dims, cosine) for Tacklebase. Content search should use the same one. Haiku is the query-understanding layer; embeddings are the retrieval layer. Two embedding models in one Supabase project would be a mistake.
• At 45 items you do not need any of this. Plain text search would do. Semantic search earns its place at hundreds or thousands of items across many suppliers, and for questions actually phrased as questions. Worth building the shape now, worth not over-buying it yet.
• Kit is extracted once at ingest, not per user. The kit in a Korda film is identical for every customer, so per-view inference pays repeatedly for the same answer. A Sonnet/Haiku pass at ingest reads the description (launch videos enumerate products against timestamps), resolves each to the catalogue, and writes content_product_edge rows for staff review. Users then get it instantly and free, and it is reviewable before it goes live.
• Deep-links use a three-tier fallback so a card never dead-ends. product resolved edge straight to the product page, collection a tag-filtered collection using the faceted filtering already live, search a query as last resort. Current mix: 25 product, 27 collection, 1 search.
• Guessed links are worse than no link. The naive fallback (infer a collection from the video's own tags) produces plausible-but-wrong destinations. That is the argument for resolving edges at ingest and saying "No kit matched" honestly when nothing resolves. 28 of 45 items carry kit; the other 17 say so plainly rather than bluffing.
• Wish Lists are not My Kit. My Kit is what you own (IDEA-002). A Wish List is what you want. Lists are per-video by default, auto-named after the video, and you can keep several. They live in the browser while logged out, with a nudge to sign in and keep them, which makes the list an acquisition hook rather than a locked door. On login the local lists migrate to the account.
• The payoff is the two lists together. Once TC knows what you own and what you want, "this video uses a hooklink you have not got" writes itself. That is the real reason to sign in, and it needs no new data, just the two lists.

Reading the POC

• Category rail is the primary navigation, per the agreed IA. Carp and Predator both populate; Sea has the two Fox Rage coastal items; Coarse and Game stay honest empty states.
• Brand filter appears now there are three suppliers. Filter to Fox Rage then toggle "Hide promo" to see the promo problem in one move: 15 items become 1.
• Try the Ask box, then open "Kit used" on a product launch and add the setup to a Wish List.
• Tags use the governed taxonomy (Technique:Zig-Fishing, Category:Lures-Soft-Plastic) so content and catalogue share one vocabulary. Note some Fox Rage tags (Category:Lures-Soft-Plastic, Category:Lure-Rods) sit outside the closed 13-value list, the same gap already logged against the Guides articles.
• Deep-link URLs are illustrative. Real ones resolve against in-stock, published products.
• Data is baked in as a static snapshot, and Wish Lists are in-memory here (localStorage in production). A scheduled Cloudflare Worker polls the feeds and writes content_item rows to Supabase (browsers cannot fetch YouTube RSS directly, CORS).