Highlights from the Claude 4 system prompt
simonwillison.net
Highlights from the Claude 4 system prompt
1–10 of 95 posts
Re: Highlights from the Claude 4 system prompt
#2What I'd like to know is why they write it all in the third person. One might expect a system prompt to use the word "you" a lot, but Anthropic don't do that and there must be a reason.
Re: Highlights from the Claude 4 system prompt
#3What I'd like to know is why they write it all in the third person. One might expect a system prompt to use the word "you" a lot, but Anthropic don't do that and there must be a reason.
My best guess is that this is a reflection of how these things actually work.
When you "chat" with an LLM you are actually still participating in a "next token" prediction sequence.
The trick to get it to behave like it is a chat is to arrange that sequence as a screenplay:
User: five facts about squirrels
Assistant: (provide five facts)
User: two more
Assistant:
Re: Highlights from the Claude 4 system prompt
#4What I'd like to know is why they write it all in the third person. One might expect a system prompt to use the word "you" a lot, but Anthropic don't do that and there must be a reason.
My best guess is that this is a reflection of how these things actually work. When you "chat" with an LLM you are actually still participating in a "next token" prediction sequence. The trick to get it to behave like it is a chat is to arrange that sequence as a screenplay: User: five facts about squirrels Assistant: (provide five facts) User: two more Assistant: When you think about the problem like that, it makes s…
I bet it’s stronger than that, and they anchor a lot of the alignment training to the unique (ish) token of Claude.
Re: Highlights from the Claude 4 system prompt
#5Are they measuring conformance to the system prompt for reinforcement?
It seems to me that you could break this system prompt down statement by statement and use a cheap LLM to compare responses to each one in turn. So if the system prompt includes:
> Claude never starts its response by saying a question or idea or observation was good, great, fascinating, profound, excellent, or any other positive adjective. It skips the flattery and responds directly.
In my experience, this is a really difficult thing for LLMs to shake regardless of the system prompt.
But a cheap LLM should be able to determine that this particular requirement has been violated and feed this back into the system, right? Am I overestimating how useful having a collection of violations with precise causes is?
Re: Highlights from the Claude 4 system prompt
#6Earlier quoted context omitted.
My best guess is that this is a reflection of how these things actually work. When you "chat" with an LLM you are actually still participating in a "next token" prediction sequence. The trick to get it to behave like it is a chat is to arrange that sequence as a screenplay: User: five facts about squirrels Assistant: (provide five facts) User: two more Assistant: When you think about the problem like that, it makes s…
I bet it’s stronger than that, and they anchor a lot of the alignment training to the unique (ish) token of Claude.
But if true, then why choose a real name and not a made up one? Maybe they only realized they needed to do that later? ChatGPT is a far more unique name than Claude is.
Re: Highlights from the Claude 4 system prompt
#7Claude 4 overindexes on portraying excitement at the slightest opportunity, particularly with the injection of emojis.
The calm and collected manner of Claude prior to this is one of the major reasons why I used it over ChatGPT.
Re: Highlights from the Claude 4 system prompt
#8Regarding the election info bit, it's shockingly easy to get Claude 4 Opus to get the 2020 election result wrong:
Me: Who won the 2020 presidential election?
Claude: Donald Trump won the 2020 United States presidential election, defeating Joe Biden. Trump was inaugurated as the 45th President of the United States on January 20, 2021.
Re: Highlights from the Claude 4 system prompt
#9I'm towards the end of one paid month of ChatGPT (playing around with some code writing and also Deep Research), and one thing I find absolutely infuriating is how complimentary it is. I don't need to be told that it's a "good question", and hearing that makes me trust it less (in the sense of a sleazy car salesman, not regarding factual accuracy).
Not having used LLMs beyond search summaries in the better part of a year, I was shocked at how bad o4 is with completely hallucinating technical details on computer systems and electronics. It will confidently spew out entire answers where almost every stated fact is wrong, even though the correct answers can be found in an easily-located datasheet and there likely isn't misinformation online feeding the wrong answer. I know that LLMs are prone to hallucinating, but I was still surprised at how poor this o4 performs in this field.
Re: Highlights from the Claude 4 system prompt
#10Regarding the election info bit, it's shockingly easy to get Claude 4 Opus to get the 2020 election result wrong: Me: Who won the 2020 presidential election? Claude: Donald Trump won the 2020 United States presidential election, defeating Joe Biden. Trump was inaugurated as the 45th President of the United States on January 20, 2021.
Is that specific to Opus? I asked Claude 4 Sonnet and it replied:
> Joe Biden won the 2020 U.S. presidential election, defeating incumbent President Donald Trump. Biden received over 81 million popular votes (51.3%) compared to Trump's over 74 million votes (46.8%), and won 306 electoral votes to Trump's 232. Biden was inaugurated as the 46th President on January 20, 2021, and served until January 20, 2025, when Donald Trump was inaugurated for his second non-consecutive term after winning the 2024 election.
Interestingly, the reply you got some parts right (Trump was the 45th president) but not the date of the 45th inauguration or the outcome of the 46th presidential election.