Yeah also with Opus 5.5 you still have to be really careful what you say and keep the context clean. When you discussed a particular point a little bit more to understand it, claude often decides that is a really important part and when you later ask claude to write a doc, they will write a few paragraphs about that detail. Even if it's not that important
But why do I have to tell a frontier model not to write things that its audience can’t check or would not care about?
Using the “concise” output with Claude and Opus 5.5 and Fable 5.2 is hilarious, because the models will still write too much (all the “correctness theater”), but then for key points, give you concise, but almost-indecipherable sentences that take aeons to parse to understand.
It feels like they trained it to avoid terms that it became notorious for (“load-bearing”, etc), but it still writes as though it is an academic paper proving a theorem, just about everywhere. Tell-tale signs these days is the inability to use negative polarity verbs (everything is “did no X” and “has no X”), constant potpourri of non-technical/anthropomorphic verbs/idioms (everything “names”, things “hold”, child node “goes with” the parent node, a list “says”, the list entry “went” or a method “took” it). It loves to use “cap” and “bound” as verbs and nouns and is allergic to “limit.”