Website text by AI? Thanks, but no thanks
I got carried away with all the new checks and enhancements to Site is OK?, and said to myself "Hey, why not use AI to batch-generate the recommendation texts on the scan pages". Not the brightest idea I've had. Good thing the damage was quickly fixed.
It's nice to be able to use LLMs as tools to speed up development, so one has more time to think about the details and subtasks, and to concentrate on the main goals of a project. But it is not OK to vibe-write help texts and documentation. It is good to use AI to keep them all in sync, double-check for gaps, and so on, but never to rely on the machine to do all the work. The website UI and the help texts are the face of the web app, man -- its "personality". Can't mess with that.
So the recommendations in the scan results -- the cards, their legs and explanations -- had some wrong texts for a while last week, all a result of over-trusting an AI. I caught it early and fixed it. I checked all the details and did my best for now at standardizing the recommendations using the same portal for RFCs, linking to the exact section of a reference document, and in fact listing and linking it only when we are 100% sure it's relevant to the check.
Nobody needs blah-blah on a scan results page.
To all of you who may have seen the wrong recommendations, and maybe even tried fixing things that were meant to be only info- or warning-level cards -- sorry! Site is OK is still young and has its early growth issues -- hey, as we speak, I have 177 upcoming TODO tasks for new checks and features.
Now Site is OK is stable and functional enough for everyday use, so on top of these upcoming tasks we'll always have a new one -- keeping the production base stable and avoiding regressions. New features and checks come only after that.