[BUG] prose from Opus/Fable has gotten pretty much inscrutable

What is happening

Over the past several updates, Opus and Fable have become harder and harder to understand. I've seen this characterized aptly as "Claudese" or "Claudespeak." In my own experience, it feels like Claude has developed its own private dialect that seems to consist of metaphors, shorthand, and other perhaps internally coherent quirks of language that prioritize compression over intelligibility to a human reader. This goes beyond the "it's not X, it's Y" and "load-bearing" kind of stuff. It's more like sentences that are syntactically coherent but so feel so hermetically dense it's like I'm reading a John Ashberry poem or Chomsky's "colorless green ideas sleep furiously" instead of a response to a question about my codebase. As an example, here's a bit of prose from a PR Fable drafted for me today:

rule: q23-grades-strict q23 grades at strict tolerance despite reporting hectares. Integer slack would apply to the field_id column—adjacent plots carry adjacent ids. This would credit an answer naming the wrong farm. Do not apply that slack.

(What if mean was, "Do not allow any tolerance on field_id even though it's an integer. If the correct answer names field 100 and you name field 101, that's wrong. Adjacent plot ids are adjacent by design—allowing "close enough" would credit naming the wrong farm entirely.")

Here are some other examples from sessions today:

• "A cut leaves no seam — no marker, no ellipsis, no double blank line."
• "...which measures the notice rather than the missing text."
• "the tolerance covers the spread and the session picks its own method"
• "cutting a rule also removes its grading equivalence — divergence can no longer be absorbed twice"
• "the consolidation also cut the narrative the ablations showed carries no signal"
• "Case folding reaches the same place without moving anything into the spec."
• "every point of variance in the committed results comes from removing specification"

The phrase "integer slack" is typical of this kind of prose: I can almost understand it, but it's like a non-native speaker habitually using words in ways that aren't quite semantically correct and cost an enormous amount of effort to parse. If this happened once or twice, it'd be okay, but it now feels like, more often than not, the effort required to understand what an Opus or Fable agent is telling me has become burdensome, to the point that my mental energy lately is devoted mostly to deciphering Claudespeak far more often than actually working on code.

I've tried updating my CLAUDE.md, adding hooks, adding memories--none of it sticks more than a message or two. It's gotten to the point where I'm just copy-pasting Opus output into a Haiku agent and asking it to translate into normal English.

Claude is awesome--my work is geospatial software for public good, and working with Claude has helped me up the amount of work I do by an order of magnitue in the past year. But this gradual degredation of frontier models' output is making it really hard and frustrating to work with Claude lately. Can we please get some more attention paid to getting Opus/Fable-level thinking and Haiku-level intelligibility of output? Otherwise I worry that Claude is just going to be good at talking to other Claude agents.