No framework, no database, no scraping — how I serve 4,380 verified local classes as static HTML
Search for a pottery class in Los Angeles and you will still land on pages that were scraped months ago, list a studio that no longer offers the session, or hide the price behind "call us." I got tired of that, so I buil
Search for a pottery class in Los Angeles and you will still land on pages that were scraped months ago, list a studio that no longer offers the session, or hide the price behind "call us." I got tired of that, so I built Local Class List: a hand-checked directory of pickleball, mahjong, dance, tennis, pottery, and flower arranging classes across 43 US cities.
As of November 2026 the site lists 4,380 classes. Every one of them was checked against the provider's own website. You book on that site, at their price. I don't run checkout, and I don't add a fee.
This post is a recap of the product rules that actually matter, and of the deliberately small stack I used to ship them.
The hard problem is not rendering pages
A local class directory dies when the data goes stale. The engineering work that matters is the set of rules that decide what is allowed to be a listing at all. They are public on the verification page, and they are stricter than a scraper would like:
- Own website only. A class is listed only if it is published on the provider's own site, and booking stays there. Classes sold only through a third-party ticketing platform are out.
- Real prices, never guessed. If the provider publishes a price, I copy it from that page. If they don't, the listing still goes up, with "See provider site" in the price slot — never an estimate. A class is marked free only when the provider actually says it is free.
- Real schedules. Dated sessions have to be in the future. Standing weekly or monthly programs are listed with their cadence, and any dates derived from a cadence are labeled as derived.
- One program, one listing. A six-week series is one listing with all of its sessions, not six duplicates.
- Local means local. A class is filed under the metro it belongs to, but the page always names the town where it actually meets. Suburbs are not relabeled as the anchor city.
Listings are re-checked on a rolling, seasonal cycle, and each class page shows the month it was last checked. When a price or date is wrong, I go back to the provider's page and fix or remove the row.
Those rules are why I did not start with a crawl. A crawler is fast at copying the open web, including the stale copies of the open web. I needed a file I was willing to put my name on.
The stack is a JSON source of truth, a Python build, and Cloudflare
There is no application server and no database in production. Each verified class lives in structured JSON. A Python build reads that file and writes static HTML: a home page, six category hubs, 43 city pages, 237 city-and-category pages, 4,380 class pages, plus a handful of utility pages (how we verify, for providers, tools, terms). The public sitemap currently has 4,677 URLs.
Cloudflare serves those files from the edge. The HTML you get for a city or class page is pre-rendered. There is no React tree, no hydration, and no origin database round-trip on a cache hit.
The only dynamic path I need at the edge is /go/{slug}. A booking button goes through that short redirect on my own domain, then 302s to the provider's page. Cache-Control is no-store on that hop, because the destination can change when a provider moves a booking URL. The redirect exists so I can count outbound clicks in aggregate and share totals with the studio they belong to. It is not a checkout.
Client-side behavior is one vanilla JavaScript file. It does four jobs:
- Remember a city preference in
localStorage, or guess the nearest metro once from a location prompt. - Power site search and the free tools from a single static index,
/assets/search-index.json— 4,380 compact records, about 2.8 MB, cached at the edge like everything else. - Hide classes that have already ended, so a search result does not keep showing a session between rebuilds.
- Fire a
provider_clickanalytics event when someone actually leaves to book.
Search also folds in a few real-world spellings: "mahjongg," "mah jongg," "pickle ball," "ikebana." That is cheaper than a search appliance, and it is honest about what the corpus is: one verified index, not the whole web.
The free tools sit on the same index. A six-question pickleball level quiz estimates a 2.0–4.0 rating and then shows matching classes. A mahjong-style quiz does the same for American, Hong Kong, and Riichi. A seasonal flower lookup starts from the month, then the workshops that use those flowers. None of these need a backend.
City-and-category pages are the useful unit
A single class page is too narrow to be the thing you send a stranger to, and a national homepage is too broad when someone wants a Saturday wheel-throwing slot. The page that has to work is the city-and-category page.
Pottery classes in Los Angeles are a good example of the rules meeting the template. The page currently lists 39 verified classes at 11 venues, with published starting prices from $52 to $350 and a median of $195 across the 32 listings that publish a price. Thirty of the 39 are open to beginners. The venues are not all "in Los Angeles" in the tourist-map sense: they also sit in Pomona, Burbank, Pasadena, Claremont, Culver City, Santa Monica, and El Segundo, and the listings say so.
The same shape holds in a smaller market. Pottery classes in Raleigh currently cover 14 verified classes at 5 venues, with published prices from $100 to $300. The flower arranging hub does the same job at category scale: 293 verified workshops and floristry classes, still checked against each studio's own site.
Every one of those pages can pre-compute the useful numbers — how many are weekend sessions, how many start after 5 p.m., what the median published price is — because the JSON already stores a real price or an explicit blank, not a guessed one. The 2026 price stats on the site are computed from the listings themselves.
What I refused to build
I refused a booking cart. If I take the payment, I become a middleman, and the listing incentive changes.
I refused to invent a price that was not on the provider's page.
I refused to treat a third-party ticket listing as a source, even when it would have been easier to fill a thin city.
I refused a front-end framework for pages that are, in the end, a title, a price, a schedule, and a button that leaves my site. Static HTML is a better default for a directory than a client-rendered table of the same data.
The cost of this approach is obvious: adding a city or a season is editorial work, not a crawl budget. The benefit is also obvious: when a page says $195 for a beginner wheel class on a Saturday in Burbank, that figure was on the studio's own site when I last checked it.
What I would do again
- Put the verification rules in public, then make the templates obey them. "See provider site" is a feature, not a hole in the schema.
- Keep one JSON file as the source of truth, and generate every page from it. Search, quizzes, and price summaries should read the same records the HTML was built from.
- Pre-render the pages people will land on from search, and keep JavaScript for preference, search, and outbound counting.
- Name the suburb. It is tempting to fold everything into the metro keyword. It is also how directories lose trust.
If you want to see the result, start at localclasslist.com. If you run a class I have missed, listings are free: email [email protected] with the page on your own site.
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.