What we learned building this
The decisions behind our products, the reasoning we used, and the mistakes that taught us something worth writing down.
Keeping the core free, on purpose
Every one of our products gives away the thing it exists to do. That is a deliberate decision with real consequences, not the absence of a business model.
How we decide what to build next
Six products is already more than a small team should carry. What gets us to say yes, and the much longer list of things that get a no.
How to close a product without wrecking the domain
Shutting something down is its own job with its own failure modes. Links, search results, the people who used it, and the data you are about to delete.
Every number a human maintains is already wrong
"10,000+ users" was true the day somebody typed it. Why the figures on this site update on their own, and the one rule that makes automated numbers trustworthy.
One queue for every music source
YouTube, Spotify, Apple Music, Deezer, SoundCloud and Tidal agree on almost nothing. What it takes to make them all behave like one queue.
What actually makes a music bot feel fast
Nobody times the gap between a command and the first note. Everybody notices it. What matters for perceived speed, and what does not.
Moderation that leaves a record
An automated action nobody can explain afterwards is worse than no action. Why every decision is logged with its reason, and what that changes for a moderation team.
The real cost of a false positive
An over-eager spam filter does not just annoy one person. It teaches a community that the moderation is unfair, and that lesson is expensive to unlearn.
Ranking a directory without freezing it
Every listing site drifts towards the same top ten. A rolling window is a small change to the ranking that changes what the directory is for.
Why we rate every episode, not every series
A single score for a twenty-four episode series throws away the thing you most wanted to know: where it was actually good.
Cataloguing anime is harder than it looks
One title has four names, three release orders and a film that is also episodes 12 to 14. Why the data behind a tracker is most of the work.
A snapshot tells you nothing
"200 players online" means opposite things depending on what happened yesterday. Why MCStat is built around history rather than the current number.
Archiving Minecraft skins before they disappear
A skin exists until it is replaced, and then it is gone from everywhere. Hundreds of thousands are kept because nobody else keeps them.
An AI that shows its working
Most AI chat hands you a confident paragraph and asks for trust. Reasoning you can read, searches you can follow and code that actually ran are a different offer.
When not to trust an AI answer
A practical list, from people who build one. The specific shapes of question where these systems are least reliable, and what to do instead.
The type check you turned off is still finding bugs
Two settings, each added for a good short-term reason, left part of a codebase unchecked. Turning them back on surfaced a bug that had already reached real data.
When a deploy changes nothing visitors can see
The fix shipped, the origin served the new version, and everybody kept getting the old one. How caching turns a successful deploy into a claim you cannot verify.
Traffic that is not an audience
Healthy-looking request numbers with almost nobody behind them. How to tell automated traffic from people, and why the distinction changes decisions.