Anthropic to release system prompts for Artifacts, latest Claude family prompts found incomplete

Last week, Anthropic released the system prompts — or the instructions for a model to follow — for its Claude family of models, but it was incomplete. Now, the company promises to release the system prompts for its newest feature, Artifacts, in the coming weeks after researchers pointed out its exclusion.

A spokesperson for Anthropic confirmed to VentureBeat that it will “add more details about our system prompts in the coming weeks, including information about Artifacts” in the next few weeks. While Artifacts, which became generally available last week, is part of the Claude family of models, the system prompts around it were not part of the latest release. Artifacts opens a window alongside a Claude chat interface to run code snippets.

In releasing the Claude System prompts, Anthropic garnered praise for its transparency from the media — including VentureBeat — as one of the few large AI companies openly giving the public a peek into how configured its models’ behaviors. However, researchers like Mohammed Sahli found the company’s claims lacking partly because of Aritifact’s system prompt exclusion.

Anthropic, however, said the reason the system prompts for Artifacts were not included in the release last week is simple. Artifacts was not generally available for all Claude users until last week. In fact, Artifacts went public only after the system’s prompt release announcement.

Why are system prompts important

AI model developers are not required to release system prompts for large language models (LLMs). However, finding these operating instructions is something of a hobby for many AI jailbreakers, and it’s almost expected the jailbroken prompts would go around developer circles after a model is released. 

But publicly releasing the system prompts opens up the LLMs more, showing how developers hope it will behave and why it will reject some user requests. 

Based on Anthropic’s system prompts documents, Claude 3.5 Sonnet, the most advanced version of its flagship model, emphasizes accuracy and brevity when answering questions. The model will not explicitly label information as sensitive or object and will avoid filler phrases or apologies. 

Claude 3 Opus, the larger model, works with a knowledge base updated as of Aug. 2023. It’s allowed to address controversial topics with a broad range of views but will avoid stereotyping and provide balanced views. The smallest version, Claude 3 Haiku, focuses on speed and does not have the same behavioral guidelines as Claude 3.5 Sonnet.

As we don’t know the system prompts for Artifacts yet, Sahli’s Medium post claims the feature is instructed to work through complex problems systematically and focuses on concise answers to queries.