One way to look at LLM’s is that they condense the vast amount of their training data and develop some sort of intuition of what pieces of information are related. To be useful, however, they need the right angle into the maze of information that they have available. In the beginning, many tips and tricks on how to correctly prompt the earlier versions of LLM models circulated on the internet a couple of years ago. But the capabilities of LLMs to find the right angle themselves advanced significantly, and what was called prompt engineering has lost its significance for everyday tasks. However, in the background, the system prompt remains a significant tool for the AI labs to adjust the behavior of their models. The system prompt is read by the LLM before the start of a new chat and is generally meant to steer the model towards useful and intended behaviour. This means that every conversation with an AI model like Claude or ChatGPT is preceded by a system prompt that the user did not write themself. So we asked, what are these system prompts about, and how did they develop?
The strongest trend over time is the ever-growing length of system prompts across providers. Some topics like Personalisation & Memory show a special growth in the system prompts across models. For GPT 5.5, one example of a system prompt snippet like this is: “User information is worth saving in memory if it is likely to change your future responses in similar situations”, which aligns with the attempts of AI providers to make their chat experience more and more personalised. Additionally, the number of tools and the necessary tool instructions that are available to the models is also only growing. If a model, for example, decides that it should update the memory, it needs to know how exactly to do so. All the instructions quickly compound, and by now, the latest system prompt of Anthropic’s Fable 5 has reached around 25 thousand words. This means that before you start a chat, an incredible amount of up to 70 pages of words (~350 words per page) have to be loaded by the model. Yes, with a context window of 1 million tokens, there is plenty of room to chat afterwards, but such long system prompts will still affect variables like the loading speed.
Looking across providers, the differences in their system prompts reflect different approaches towards LLM deployment. Some model providers like SpaceXAI with their Grok models prioritise speed and keep the system prompts to a minimum, while other providers like Anthropic and OpenAI try to make their models useful by loading more tools and tool instructions into the default context of a chat. Beyond simple usefulness, Anthropic’s Claude models stand out with the highest share of behavioural guidance. Some of it is about personalisation and copyright, but a significant share is also about safety and user wellbeing – especially compared to other model series. In a way, the system prompt is only the last step after the whole training pipeline that influences the behaviour of AI models, and there are other ways (like reinforcement learning) that can make a model safer and more reliable. But it still is an interesting signal on what AI labs deem important and guideline-worthy.
Note: Extracting the system prompts has become increasingly hard over time and has to be treated with a degree of caution, but you can find a great collection here on GitHub. The system prompts from there were split into paragraphs and LLM-classified into selected topics.