Writing
Building with AI, judging whether it works, and the ideas underneath — including honest write-ups of what did not. Subscribe by RSS to get new posts.
Subscribe by RSSA 2,500-year-old Sanskrit grammar resolves rule conflicts with a single metarule stated in three words. Most system prompts I have read resolve theirs with a line in caps, added under pressure, that nobody has gone back to check. On what the older engine got right.
Traditional software tests assume one correct answer. AI systems do not work that way, which is why so many teams fall back on testing by hand and calling it quality. A practical case for building a repeatable evaluation process, and what it takes to run one with every important change.
Regular search only looks at file names — it never looks at what is inside a file. A build write-up of an AI search tool that reads the file itself, why I deliberately avoided a generative model, and how I made sure every result points at something real.