Cards are for scanning. These are the writeups — what the problem actually was, which decisions were load-bearing, and how each system is proved to work rather than asserted to.
For five years all my tests used one simple rule. If you run a test two times you get the same answer. Coding agents do not follow that rule. Most of what I know as a tester still works but first I had to change what a pass means.
A landing page that makes the pitch by letting you do it. Shway is “Tinder for your plans”, so the hero is a deck of plan cards you can drag, flick and match with — the product's core gesture, playable before the download.
Splitwise got worse and it never understood UPI. Squared Up is the version I wanted. The interesting part is not the app — it is that the money math is a framework-free Python package, proved out before Django ever sees it.
An LLM that says “BUY” is worthless. An LLM that says BUY at 0.78 confidence, with a 12-month target, two named reasons and two named risks — and gets retried when it fails to produce them — is at least something you can argue with.
A documentation assistant that would rather say nothing than guess. Every answer is retrieved from GitLab's own docs, cited back to a source URL, and refused outright when retrieval cannot clear a relevance floor.
A product, not a repository. Steno is a writing standard that stops AI prose announcing itself — sold as a one-time purchase, with the rules, the copy, the brand, the landing page and the payment flow all built by the person who wrote the rules.