Both Claude Opus 5 and Fable 5 have a common problem: they speak in riddles and metaphors. Something is hanging off somewhere, rides on something, lands somewhere, or is dare I say load-bearing. I call these "Claude-isms." Besides being annoying, the real problem is that it forces the user to spend time and energy "decoding" the meaning, since the actual literal meaning is being buried under made up coinages, unnecessary metaphors and endlessly figurative language.
Anthropic is aware that something is not right, and have released a concise output style for Claude Code, but fundamentally the problem isn't verbosity it's compression and spread of the literal meaning of what it's trying to say.
There's decades of research around what makes for easy to understand language for humans, and in my experience Claude seems to respond in ways that go against most of the established literature around this area. Its failures broadly break down into 5 categories as I see it.
Main Failure Modes
Language Complexity
Using language that is too complex for no good reason. Diverging, rather than different from. Bydirectionally, rather than both ways. Its own made up coinages about problems. The two-way model about something or other. This is just more complex to sound more sophisticated.
Language Compression
This is the big one. The metaphors, the refusal to be literal about anything. The flowery embelishments it uses to talk about basic things, that's your load-bearing, riding on, hanging from, and so on. It forces you to translate the response into something you can actually map in your brain to something logical. If it's load-bearing, it breaks without it. If it's riding on, it's dependent on something.
But the problem isn't just using metaphors and fancy language, you can have a perfectly legible piece of text that is still somehow very difficult to understand. Let's look at this example:
my second paragraph answered with a diagnosis instead of stopping at the definition
In order to understand what this is talking about, I have to refer back to the second paragraph. Then I need to find this diagnosis, and then figure out what it means by its definition. There's significant amount of cognitive load put on me as the reader here, I have to piece together a bunch of context. Instead of it just telling me directly what the diagnosis was, why it was wrong, and what it should have done instead.
Saying instead of doing
You're absolutely right!, I've read all of it now, both halves., The twist that gives the paper its name..., Let me look into that now.
It likes to pad out its responses with something that looks balanced and nice, but at the cost of the users attention. The user must parse this in order to get to the real information they care about. It wastes time, and tokens for that matter. If you're going to do something, just do it, and tell me what actually requires my attention.
Formatting
This is another critical one. It tends to spit out overly formatted soup of different highlighting styles, lists, and tables, without considering what would actually be the most effective tool for communicating that particular information. More often than not, simple prose with minimal highlighting would do a much better job that a list that contains prose, file references, bolded/highlighted text, etc. It lacks a proper reasoning step for figuring out the best way to visualize a response.
Not aligned to user proficiency
This is pretty basic. It doesn't have a running memory of the users proficiency in the subject matter, so it either ends up explaining things with terminilogy the user does not understand or in too simple terms that end up being confusing to a professional. If the user is proficient in the subject, using overly plain language will also make it more difficult to understand the response.
What is the root cause?
Without knowing more about the way the model is trained, we have to speculate here, but seeing as no amount of output style massaging or CLAUDE.md notes seem to have a significant impact, it's most likely down to the post training process Anthropic uses to train the model. How they classify and rate responses during that process is what ultimately will determine how it behaves. Anthropic themselves say that they use Claude to rate responses during post training to hone in on the style and character of Claude. If it's told to prefer figurative, metaphorical, witty responses, it will keep favouring that style of speaking and it will sound more and more like itself over time. This is what I think has happened with the latest models. These Claude-isms have been an issue before already, but it's clearly gotten to a breaking point recently.
It seems pretty evident that there is no weight given to clarity and the ease of understanding a response. And to be fair, that is a very difficult metric to establish. What makes for easy to understand language isn't as simple as being less verbose, or using fewer complex words. How context is spread out and how much a reader has to hold in their head while reading a sentence plays a very significant role and that isn't a simple metric to derive in a mathamatical way.
I tried to use some existing tools used for language research, and in my limited testing it was clear that deriving an aggregate "clarity" score is not a simple problem to solve, so I do understand the position Anthropic is in. However, at least from the outside it doesn't seem like they're concerned enough about the "user experience of language" nearly enough, and have put way too much stock into the appearance of character and cleverness.
What can you do now?
Anthropic has released a concise output style, which can help a bit and they also say it's a temporary measure, but in my testing I found that the training has weighted the model so heavily towards this style that it's not really effective. And fundementally, the issue is not being concise, it's how compressed the language is and how spread out the context is. In other words: how much effort does it require to get to the literal meaning of what it's saying. This must be fixed at the model level.
The cost is clear: it's harder to get work done with Claude. We're in an era now where people are running multiple agents doing all kinds of things in parallel, so the bottleneck is the human. The longer and the more effort it takes to decode what the agent is saying, the worse off we are. Our attention or "cognitive load" is ultimatly also a finite resource, wasting it on just trying to understand the agent isn't ideal.
Top comments (0)