I've been hearing a lot about the Claude system prompt and I'm shocked at how massive it is. A token calculator claims it's over 24,000 tokens! I understand prompt caching exists, but it seems really inefficient to process this many tokens for each query. For instance, there's about 1,000 tokens dedicated to discussing CSV files—why include that in queries that have nothing to do with CSVs? Can someone clarify if I'm missing something? Is there a way to disable this in the Claude interface?
5 Answers
I noticed the same issue! Even a slight increase, like enabling artifacts, reduced Claude 3.5's accuracy significantly, and that was a minor prompt addition compared to the full 24K. Just goes to show that longer prompts can be problematic.
The official Claude system prompts are published and they look much shorter than 24K! But even the latest ones still exceed 2,500 tokens for Sonnet 3.7. Here’s a link to the official documentation if you're interested: [Claude System Prompts](https://docs.anthropic.com/en/release-notes/system-prompts#feb-24th-2025).
Having such a long system prompt isn't just a hassle because of rate limits; it can also hurt the performance of the model. The context length can make it harder for the system to generate relevant and accurate responses. So yeah, it’s something to be cautious about.
Honestly, these systems seem unsustainable with how large these prompts are getting. It’s just too much to handle.
This is an absolute mess! They really need to streamline their approach.
True, but keep in mind those additional tokens also cover how the system should utilize tools. Not a total waste!