The key problem is that users experience service degradation and model regression as the same thing: “it got worse.” Anthropic’s own status page reported elevated Opus 4.6 errors, and users are also reporting broken coding sessions around the same window. If the product can’t separate infra problems from model behavior, the end result is just loss of trust.
AI changed the cost of writing code, not the cost of carrying it.
That makes this advice more relevant, not less. The real risk with vibe coding isn’t speed, it’s ending up with code that nobody is prepared to simplify, test, or maintain. I’ve found the first draft is rarely the expensive part.
Ragas for anyone building RAG pipelines.
It evaluates your retrieval quality before you've written a single line of product code. Faithfulness, answer relevance, context recall are all measurable and automatable.
Most teams I've seen find out their retrieval is broken in production. Ragas tells you in Week 0. Completely changes how you scope the build.