How to Build Reliable AI Products: A Production Engineering Guide
AI demos are easy. Reliable AI products are an engineering problem.
Technische Posts, Produktentscheidungen und ehrliche Beobachtungen, geschrieben beim Bauen, nicht danach.
The best notes if you want to understand how I think, build, and ship.
AI demos are easy. Reliable AI products are an engineering problem.
MV3 is four moving parts, manifest, service worker, content scripts, and UI, not a single script tag anymore.
Stop presenting projects. Start shipping a system that compounds trust.
Build in public
Was shipped wurde, was kaputtging und was sich änderte, geschrieben, als es passierte.
Shipped a popup refresh focused on first-glance clarity: the popup now states the job plainly, the current-tab shortcut says exactly what it does, and the empty state lost its duplicate copy. Small release, but every session starts at the popup.
Rebuilt the Amethyst popup around one question: where did today go? It now opens with the daily total, top sites, and a single button into the full dashboard. Removed the tracking badge, comparison chip, sync button, and feedback card, the glance view answers fast, then gets out of the way.
A run of notes on making this site readable by AI agents: the MCP server, the public API, the data model, and the security lessons from exposing it all. Personal sites are becoming APIs, this series documents doing it for real.
Wie sich KI-Funktionen verhalten, wenn echte Menschen sie nutzen, nicht in kontrollierten Demos.
You fine-tune the voice, not the perception. Perception is a commodity you rent; voice is the thing you cannot buy.
The expensive failures were not training failures. They were infrastructure failures that produced plausible output instead of an error.
Praktische Guides für Manifest V3, Browser-Tools und Extension-Distribution.
The Web Store review is a product gate, privacy policy, assets, and MV3 compliance matter as much as your code.
Real products need friction by design, ResistGate makes you type before distraction, not just block URLs.
Preise, Produktentscheidungen und was ich gelernt habe, als ich Dinge bauen wollte, die Menschen wirklich kaufen.
Game monetization splits cleanly into predictable-and-fair and random-and-manipulative. Regulators are banning the second, and the same line runs through how any product charges.
Retention won through dark patterns is debt. It builds regulatory and reputational liability and makes every future price increase impossible.
Architektur-Trade-offs, Primitive statt Komplexität und dauerhafte Systemmuster.
Predict locally, decide centrally, and make rejection a first-class state.
The transport moves messages. Your protocol decides whether the product stays coherent.
Wie ich Entscheidungen treffe, fokussiert bleibe und ohne Team vorankomme.
You can't skip the ten. Every million starts as ten people who actually stayed.
The right model is the one your product's costs, users, and habits actually point to, not the trendy one. Ten questions get you there.
Schlaf, Routinen, Lernen und was mich wirklich auf voller Kapazität hält.
Notizen nach Thema und Monat filtern.
Thema
Monat