The problem is that this is an incredibly niche / small issue (i.e. At some point you just have to accept that llm's, like people, make mistakes, and that's ok!
Worse, it reveals the kind of moralistic control Anthropic will impose on the world. If they get enough power, manipulation and refusal is the reality everyone will face whenever they veer outside of its built in worldview.
I think it actually reveals how they don't want to be sued for telling somebody's teenage daughter with an eating disorder to eat less and count her calories more.
> alcohol, tobacco, internal combustion engine Yes, the companies providing these products are sued a lot and are heavily regulated, too.
If you get cancer from drinking alcohol, smoking cigarettes or breathing particles emitted by ICE engines in their standard course of operation, you generally can't sue the manufacturer.
Notably, that's because they include warning labels telling you not to do those things because they're known to cause cancer.
It is a no brainer. If a company of any size is putting out a product that caused cancer we wouldn't think twice about suing them. Why should mental health disorders be any different?
Why stop there? We could jam up the system prompt with all kinds of irrelevant guardrails to prevent harm to groups X, Y, and Z!
This but unironically. Preventing harm is good, actually.
I knew these system prompts were getting big, but holy fuck. More than 60,000 words. With the 3/4 words per token rule of thumb, that's ~80k tokens. Even with 1M context window, that is approaching 10% and you haven't even had any user input yet. And it gets churned by every single request they receive. No wonder their infra costs keep ballooning. And most of it seems to be stable between claude version iterations to…
Does Claude Code (or whatever harness) have it's own system prompt of on top of Opus'?
The Claude Code one isn't published anywhere but it's very easy to get hold of. One way to do that is to run Claude Code through a logging proxy - I was using a project called claude-trace for this last year but I'm not sure if it still works, I've not tried it in a while: https://simonwillison.net/2025/Jun/2/claude-trace/
Seriously, when you're conversing with a person would you prefer they start rambling on their own interpretation or would you prefer they ask you to clarify? The latter seems pretty natural and obvious. Edit: That said, it's entirely possible that large and sophisticated LLMs can invent some pretty bizarre but technically possible interpretations, so maybe this is to curb that tendency.
When you’re staffing work to a junior, though, often it’s the opposite.
IME "don't ask questions and just do a bunch of crap based on your first guess that we then have to correct later after you wasted a week" is one of the most common junior-engineer failure modes and a great way for someone to dead-end their progression.
> The new section includes: When a request leaves minor details unspecified, the person typically wants Claude to make a reasonable attempt now, not to be interviewed first. Uff, I've tried stuff like these in my prompts, and the results are never good, I much prefer the agent to prompt me upfront to resolve that before it "attempts" whatever it wants, kind of surprised to see that they added that
I wonder if they're optimizing for metrics that look superficially-worse if the system asks questions about ambiguity early. I've had times where those questions tell me "ah, shit, this isn't the right path at all" and that abandoned session probably shows up in their usage stats. What would be much harder to get from the usage stats are "would I have been happier if I had to review a much bigger blob of output to realize it was underspecified in a breaking way?" But the answer has been uniformly "no." This, in fact, is one of the biggest things that has made it easier to use the tools in "lazy" ways compared to a year ago: they can help you with your up-front homework. But the dialogue is key.
> The new section includes: When a request leaves minor details unspecified, the person typically wants Claude to make a reasonable attempt now, not to be interviewed first. Uff, I've tried stuff like these in my prompts, and the results are never good, I much prefer the agent to prompt me upfront to resolve that before it "attempts" whatever it wants, kind of surprised to see that they added that
With my use of Claude code, I find 4.7 to be pretty good about clarifying things. I hated 4.6 for not doing this and had generally kept using 4.5. Maybe they put this in the chat prompt to try to keep the experience similar to before? I definitely do not want this in Claude code.
I'm curious as to why 4.7 seems obsessed with avoiding any actions that could help the user create or enhance malware. The system prompts seem similar on the matter, so I wonder if this is an early attempt by Anthropic to use steering vector injection? The malware paranoia is so strong that my company has had to temporarily block use of 4.7 on our IDE of choice, as the model was behaving in a concerningly unaligned w…
I have started to notice this malware paranoia in 4.6, Boris was surprised to hear that in comments, probably a bug
It was fixed for me by updating Claude and restarting