The build was the easy part.

Google Maps will tell you two things reliably: there's a hall nearby, and yes, it's open right now. Ask it anything more useful, what a day pass costs, whether there's a café worth sitting in afterward, whether the place is even bouldering and not a rope hall that happens to show up in the same search, and it shrugs. Someone standing in a city they don't live in, phone in hand, wants five answers: is there a hall here, is it open, what's a ticket, is there somewhere to sit afterward, can I rent shoes. Every decision that follows in this directory, from what gets filtered to how long a description is allowed to run, traces back to that list. The gap between what Maps answers and what people are looking for is the whole reason for the directory to exist.

Every gym starts as a scrape: name, address, opening hours, and however many reviews are public, pulled in from the web rather than typed in by hand. Reviewer names and any other identifying details get stripped out before anything is stored, only the review text and star rating survive. That raw pull isn't usable on its own. It becomes an actual listing through an enrichment pipeline, built together with Claude over the months that followed, that reads through the scraped reviews and the hall's own website and decides what to fill in: which amenities are genuinely present, what the real opening hours are, whether the tone of the reviews suggests a hall worth trusting.

Every amenity tag a gym shows now, café, kids' area, training area, and the rest, comes out the other end of that pipeline. None of it is typed in by hand, and none of it existed at all in the first version. Whenever the pipeline got something wrong, the fix was never just that one hall, it was a rule specific enough that the next hall with the same problem wouldn't need a human to catch it again. What that pipeline gets right and wrong, and how those rules built up over six months, is most of what the rest of this is about.

Airtable was the obvious pick in February. I already knew it, it keeps data clean, and it's fast to wire up when you want to be building by the weekend, not reading documentation. I knew going in about the 1,000 API calls a month on the free tier. What I didn't know was how often Claude Code would hit it during builds, since I'd never paired the two before and had no feel yet for the call volume that generates. The site blew straight through the limit in April. Once that happened, I found the fix: stop calling Airtable at build time at all. A local cache, and a build that reads from a checked-in src/data/gyms.json instead of hitting the API live. Airtable wasn't a bad call, it just needed a build pattern I hadn't worked out yet. One smaller lesson lived in the same table: after some experimenting with field types, leaving cities as free text instead of a locked dropdown, so new ones could get added straight from the crawl, turned out to be the better call.

Leaflet over Google Maps was Claude's suggestion, picked because it was easier to wire up. Google Maps wouldn't have been impossible to make GDPR-compliant, just a lot more engineering to defer properly. Leaflet paired with OpenStreetMap tiles made it easy, almost by accident: OSM tiles are just static image requests, trivial to hold back until someone consents. That's its own section further down.

Astro's static generation means every page gets built ahead of time, and that build step is also when a handful of scripts run to check that everything's still in order: one confirms the hall counts in the visible copy match the actual data, another pings roughly 200 gym websites looking for dead links, a third flags any closure banner that's overdue for a recheck. Catch a problem there and the build fails before anything ships. None of that works if pages were rendering on the fly, there'd be no fixed moment to run the checks in.

Years of WordPress and Elementor had taught me a way to build websites, but it wasn't suitable for what I wanted to do here without dragging in a lot of bloat. Boulder gyms weren't the first idea either. I tried out a handful of directory categories before settling on this one, weighing search volume, how much decent competition was already out there, and how manageable a first AI-built directory needed to be. Boulder gyms in Germany hit the right size on all three.

I didn't just open Claude Code and start typing at it. Requirements came first, worked out in a plain chat conversation: what the directory needed, how the data should be shaped, what mattered for search. Then five prompts, each one a slice of that document, fed to Claude Code one at a time. By the end of the weekend I had a functioning MVP: filters, a map, sorting, a first attempt at enriching the data.

Of course the traffic didn't show up in the first week, that would have been a lot to ask. I kept working on it anyway for a while, mostly because I wanted to refine what was there, not because anything was telling me to yet. The real signal took about two months: Google started surfacing pages on its own, traffic began climbing, and that lined up with a handful of blog articles I'd been writing alongside the directory that were pulling in search traffic of their own. That combination, not any single moment, was enough to say this was worth continuing rather than becoming another abandoned side project. It helped that continuing cost basically nothing: hosting was already in place from years of building other sites, so there was no infrastructure bill making the decision for me either way.

Part of that first stretch was just watching. Which cities started ranking on their own, without any push from me, was the clearest signal for where real demand sat. The blog posts helped in a second way too: a handful of them picked up backlinks from other sites, which did more for the domain's authority than anything on the directory pages themselves could.

What happened in between: the weekend produced an MVP, lo-fi and “a little rough,” and once I saw I could build on it, I built out more cities first. At some point the lo-fi look stopped being charming and started getting in the way, so it went from barely-styled to at least good enough not to distract from the content.

Germany has hundreds of climbing venues, and a fair number of them will argue with you about whether they count as a boulder gym. Rope halls, DAV club centres, high ropes courses, outdoor crags, fitness studios with one lonely wall, university halls, kids' formats, and one memorable venue that projects video games onto the holds, so you can technically boulder a level of Mario. Working out which of those belong in a boulder directory is the actual product decision of the whole project, and it took four tries to land on a rule that held.

A directory's inclusion criteria are the product. They can't be derived from data, only argued into existence and written down somewhere. A handful of halls still just need manual checking, there's no way around that, but every one of them also teaches the rule something, and it gets a little sharper each time.

The amenity data used to render exactly the same wherever it showed up, a card and a full profile got identical treatment, which in hindsight was a bit lazy. It now comes from one shared component and looks different in each place on purpose: compact pill chips on a card, a two-column labelled list on a profile. On a card you're comparing halls and scanning for one thing, so density wins. On a profile you've already picked the hall and you're reading, so legibility wins. Same data, dressed for two different occasions.

The original insight still holds up: when a review complains the café is expensive, that's not bad news for the café amenity, it's confirmation the café exists. Nobody complains about the price of a thing that isn't there. Negative sentiment is still evidence.

What six months added is the harder half, and it's basically the opposite claim: silence doesn't mean absence, and treating it like it does will quietly wreck your data. A sweep that proposed removing amenities no review had mentioned was wrong 8 times out of 10. One hall's outdoor area got flagged for deletion across 23 silent reviews, until digging in showed a single reviewer had described it in loving detail using the word „Außenfläche“ instead of the „Außenbereich“ the pattern was searching for. A café flagged the same way was disproved by one look at the hall's own website, which, in fairness, is where most people would have checked first. And a kids' area that looked solidly confirmed by four separate reviews was actually four phrases inside a single review, written by the one person arguing the hall didn't have one.

Reviews were never the only input either. Website crawls filled in plenty that reviews left silent, and when even that was unclear, the only way through was checking the hall's site myself and turning whatever I found into a rule for the next hall like it.

An AI can tell you what a review says. It genuinely cannot tell you what a review's silence means, and treating the two as equivalent is how a directory gets quietly less accurate while looking more thorough the whole time.

222 of 226 gyms have an individual description now, and every single one has to survive the same test: swap the hall's name for any other hall's, and if the paragraph still reads fine, it says nothing and gets rewritten. Getting there took more than a one-line prompt. Claude Code writes from a set of guidelines I've built up over the months since: rules for tone, for what's worth mentioning versus left out, and specifically for how to handle a negative review, fold it in plainly rather than smoothing it into something vague. Those guidelines get revised the same way everything else here does: whenever one produces a description that reads flat or misleading, the fix becomes the next rule. Nothing gets quoted from a review verbatim either, partly for copyright, mostly because it's someone else's sentence, not mine to sell.

The detail I genuinely didn't expect to lose an afternoon to was German grammatical gender in front of a business name. It's das Stuntwerk Essen, but die CAMPUS Boulderhalle and der Boulderwürfel, and there's no shortcut, no rule of thumb, nothing. A dedicated module resolves the article per hall, and where it can't tell, it rewrites the sentence to avoid needing one rather than guessing and hoping. Getting this wrong across 244 pages is a tiny error that reads as a huge one: it's exactly the kind of slip that makes a site feel like nobody wrote it.

Descriptions weren't the only content decision. Every city page opens with its own intro paragraph, built from the specific halls in that town rather than one template with the city name swapped in, and a hall's district gets mentioned naturally wherever it's genuinely part of the story, not stuffed in for the keyword. Both city pages and individual gym profiles carry an FAQ section at the bottom: real questions someone would ask before a first visit, opening hours, kids, renting shoes, answered plainly. They're there for the visitor first, but they also happen to be exactly the kind of clearly labelled, quotable content that search engines and AI answer engines are built to pull from. Blog posts get headlines that say what they're about rather than something clever, and every gym profile's H2s carry both the relevant keyword and the hall's own name, so a page has a shot at ranking for someone searching the hall by name and not just the category.

The map on the site won't load until a visitor says yes to it, and that's not a blank grey rectangle with a cookie banner stapled on awkwardly. It's a designed state with its own layout, explaining what loading the map means before it happens rather than after. Fonts are self-hosted instead of pulled from a CDN, because loading them from Google would hand every visitor's IP address over before anyone had agreed to anything at all. Analytics fires after consent or not at all, and the banner offers accept and reject with equal visual weight, not a big green button next to a small grey link that's clearly hoping you won't notice it.

The consent gate is a page state, the same as an empty state or an error state, and most sites design it like they forgot it was there. This one didn't get to forget.

Each version had a specific job, not just a fresher coat of paint. V1 in February existed to prove the product worked at all, functional and technically sound, nothing more. Once that was settled, usability became the driver behind the next two versions. An earlier redesign focused on making the site easier to use while the technical foundation stayed untouched, and started laying groundwork for monetization to see what would stick. This latest redesign continues that thread: another usability pass, and also where the richer profile type gyms will eventually pay to upgrade to, shown below, got built out for the first time.

The redesign shipped in August, six months in, onto a site that already had real, indexed pages doing real work in Google. That changes what a redesign is allowed to do: it can't move a URL, full stop. The sitemap was checked before and after against the same 385 addresses, and nothing needed a redirect. A directory's only real asset is the pages Google already trusts, and a redesign that reshuffles URLs for a fresher coat of paint is setting fire to that trust to make things look nicer.

The comps themselves came from a separate design tool, handed over as flat HTML. A few got built differently than drawn, on purpose, and the reasoning for each is written down in the repo rather than quietly lost. That's the normal, slightly untidy gap between what gets designed and what ships, and it's usually the part a portfolio leaves out because it's less flattering than a clean before-and-after.

This started as a practice project. Six months in, it looks more like something that could genuinely work as a business, and the plan has shifted accordingly: keep expanding coverage, keep earning rankings with real content, and start turning that traffic into revenue rather than just watching the numbers climb. Roughly 75 German cities above 50,000 people still haven't been checked for a hall, and that's still the most straightforward way to grow.

The monetization side is still early, but it's not vague. Premium profiles are the first piece: gym owners can already claim their listing, and a paid tier on top of that adds photos, a fuller description, and better placement. After that, the plan is affiliate links where they make sense, paid placement in the guide articles and city pages themselves, and direct brand cooperations with gyms or gear companies who want in front of this specific audience. None of that works without the content and the rankings coming first, which is exactly why that part came first.

Not every gym will pay for a premium profile, so there's a second track running alongside it: a badge gyms can embed on their own site, just released, that links back to their directory profile. It costs a gym nothing to add, and it hands the directory the one thing rankings depend on, backlinks, without needing a single sales conversation.

The build is the cheap part. Budget for the judgment, not the code.

Any rule that runs on its own needs a measured error rate before you trust it blindly. The silence rule was wrong 8 times in 10 and felt completely right the entire time it was running, which is the scary part.

As of early August, before the individual gym profiles rolled out in full, July sat at roughly 50,000 impressions and a few hundred visitors. Enough to keep going. What a much larger profile-page expansion does to those numbers is the next real test.

The original point still holds, and it's aged well: AI removes the roadblocks that stop you from acting on your thinking, but it doesn't do the thinking for you.