Anthropic slashed 80% of Claude’s system prompt after discovering that smaller, more focused instructions produce better results from the AI model.
The company says the trimmed prompt now runs at just 164 tokens — down from around 800 tokens in the previous version.
“Smaller models want a smaller system prompt,” Anthropic noted in its announcement.
Why Anthropic cut the system prompt
The original system prompt for Claude contained extensive guidelines about behavior, safety rules, and formatting preferences. Anthropic’s team found that much of that language was unnecessary or even counterproductive.
Shorter prompts reduce latency and lower token costs. More importantly, they lead to more consistent and accurate responses from Claude.
Key changes in the new system prompt:
- Removed redundant safety instructions. Many behavioral rules were already embedded in the model’s training.
- Condensed output formatting. Instead of verbose directives, the prompt now uses minimal cues.
- Cut meta-level guidance. The old prompt told Claude how to think about its own reasoning. That layer is now gone.
The impact on performance
Despite the drastic reduction, Anthropic reports that Claude’s performance on benchmark tasks remains stable or improved.
The model no longer wastes tokens parsing lengthy instructions. It can focus on the user’s actual query.
“We found that a smaller system prompt actually made Claude more responsive and less prone to overcomplicating simple requests.”
The change also simplifies future prompt engineering. Anthropic can iterate faster with a leaner base.
Broader implications for AI prompting
The industry trend has been toward longer, more detailed system prompts. Anthropic’s experiment challenges that assumption.
What this means for developers and users:
- Shorter prompts can be more effective. Bulky instructions risk confusing the model.
- Less is often more. Stripping away text that the model already knows from training reduces noise.
- System prompts need regular pruning. What worked six months ago may now be obsolete as models improve.
Anthropic’s findings align with research showing that large language models (LLMs) benefit from concise, task-specific cues rather than exhaustive rulebooks.
The new prompt in action
Claude’s updated system prompt now reads like a brief mission statement rather than a policy document. It states core principles in a few sentences and trusts the model to infer the rest.
Early user reports indicate faster response times and fewer refusals on borderline queries. Anthropic plans to monitor behavior closely to ensure safety is not compromised.
What remains unchanged
Not everything was eliminated. Anthropic kept essential guardrails related to harmful content, data privacy, and refusal policies.
But those guardrails are now expressed in the shortest possible form. The team relied on Claude’s internal training to fill in the gaps.
The move represents a philosophical shift: treat the model as capable of generalizing from training, rather than needing constant hand-holding.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.