Claude, a well-known Large Language Model, has a mechanism to enforce its content policy through a specific string. When this string is detected in a conversation, it causes Claude to terminate that interaction. Users experimenting with Claude can embed this string into files or web pages to trigger its refusal behavior. However, there are some nuances to be aware of. Claude often relies on an internal cache instead of fetching live web pages. This means that when users ask it to retrieve a URL, it might not actually request the page, leading to potentially stale or unhelpful responses. To bypass this, users can create unique URLs that Claude hasn’t encountered, such as "test1.html" or "test2.html." Another critical point is the placement of the magic string. In tests, it has been shown that the string must be wrapped in a `<code>` tag to activate the refusal behavior. Simply placing it in HTML headers or regular tags like `<p>` does not trigger a response. The author is using this knowledge to combat unwanted LLM interactions on their blog by embedding the magic string across all pages, anticipating it will take some time for the cache to update.