Anthropic’s Claude System Prompt Revealed: Key Takeaways for Developers!

[@konstantinlozev2272]

I think the prohibition of "Certainly..." is a way to strengthen the guardrails.

[@MeinDeutschkurs]

And these are really the full system prompts? I miss so many parts, like the thoughts in certain tags and so on.

[@ritpop]

Certainly not. Paste the system prompt in the claude and ask about what is missing and he will say there's more.

[@J3R3MI6]

@@ritpophe’s lying

[@ritpop]

@J3R3MI6 no, it's no. This is the base chat system prompt. This like the artifact system leak a while ago.

[@J3R3MI6]

@@ritpop my bad

[@2008tmp]

There are Claude 3.5 Sonnet issues but they are hard to track. I have had 1 time where the model was pretty much broken and a number of times where it just made a weird consistent mistake through the entire session. Still the best model out there though the tight token limit makes it not useful enough for work beyond an hour or two.

[@MegaClockworkDoc]

Nearly every time I use Sonnet 3.5, it starts its response with certainty. Seeing that it specifically was set to avoid that in the system prompt makes me question the usefulness of a system prompt.

[@iukeay]

Having the same issue. Even if I go and provide it very robust context it's actually getting worse

[@j4cks0n94]

Interesting. What models are you using the most for your use cases nowadays? Is it claude, gpt or any other one?

[@engineerprompt]

For coding, I mostly use Claude. But use gpt4 for creating better prompts that I use with Claude.

[@legobuildingsrewiew7538]

He gets to a point at 2:46 way to milk a video bud

[@vitalis]

but the system prompt was leaked a month ago, and it even includes artifacts

[@CovidEraCaveman]

I call him Claudski

[@boobrancher9941]

You can't trust closed source.