AI Card Bao BCAST.AICARDBAO.COM

Industry News

ChatGPT and the set of rules that Claude read before speaking have been accumulated into a warehouse of 72,000 stars.

Before you even utter your first sentence, the chatbot has read a long instruction manual. That is the system prompt word written by the manufacturer, and users cannot see it under normal circumstances.

There is a repository called System Prompts Leaks on GitHub that collects these things. It currently has more than 72,000 stars.

The coverage is very wide: ChatGPT, Claude, Gemini, Grok, Cursor, Perplexity, DeepSeek, plus dozens of other products. There is one folder for each company, and the earliest version can be traced back to May 2025, updated every day. The sources are either leaked or made public by the authorities themselves.

The items inside are quite interesting to read.

Claude's prompt states that he "never swears unless asked to do so or unless the other party swears profusely." List format is not allowed when rejecting users, the reason is that "a little more effort will help soften the blow."

ChatGPT's August advisory asked it to avoid "unwarranted or flattering compliments."

The length is also amazing. That single file of Claude Fable 5.2 alone has about 67,000 English words, the length of a novella. At the beginning of each conversation, the model has to go through these first.

Developer Simon Willison commented that this is the most complete collection of leaked system prompt words he has ever seen.

This warehouse is useful for three types of people.

People who write prompt words can directly see how the big manufacturers write them. Many of the "prompt word techniques" that are circulated are guesswork. Here is the original: how to draw boundaries, how to describe tone, and how to handle exceptions were all developed after a lot of evaluation.

People who make products can see why competing products behave the way they do. If a certain model suddenly becomes verbose or cautious, it is often more accurate to look at the difference in prompt words between the two versions than to guess that the model has been updated.

Ordinary users can understand one thing: a large part of the "model personality" you think is written, not trained.

As a reminder, the accuracy of the leaked version cannot be verified item by item, and the model may complete the content out of thin air when being lied to. It's okay to use it as a reference, but it's not appropriate to quote it as an official document.

Related