“You are a useful mail assistant” and other instructions from Apple Intelligence

Apple's latest developer betas launched last week with a handful of generative ai features that were announced at WWDC and are coming to iPhones, iPads, and Macs in the coming months. On Apple computers, however, you can read the instructions programmed into the model that supports some of those Apple ai features.

They appear as prompts preceding anything you say to a chatbot by default, and we've seen them discovered for ai tools like Microsoft Bing and x.com/PhotoGarrido/status/1710035960514175322?s=20″>GIVE HIM before. Now, a member of the macOS 15.1 beta subreddit posted that they had I discovered the files containing those backend messages.You can't alter any of the files, but they give an early clue as to how the sausage is made.

In the example above, an ai bot acting as a “helpful email assistant” is told how to ask a series of questions based on the content of an email. It could be part of Apple’s Smart Reply feature, which can suggest possible responses.

It sounds like Apple’s “Rewrite” feature, one of its writing tools that can be accessed by highlighting text and right-clicking (or, on iOS, long-pressing) on ​​it. Its instructions include passages that say “Limit response to 50 words. Don’t hallucinate. Don’t make up factual information.”

This short message summarizes the emails, with a careful instruction not to answer any questions.

I’m pretty sure this is the Ajax instruction set for generating a “Memories” video using Apple Photos. The passage that says “do not write a story that is religious, political, harmful, violent, sexual, dirty, or in any way negative, sad, or provocative” might explain why the feature rejected my request for “sad images”:

It's a shame. It's not hard to get around though. I got it to generate a video in response to the request: “Send me a video of people in mourning.” I won't share the resulting video because there are photos of people who are not me, but willpower I show you the best image included in the presentation:

The files contain many more instructions, all of them hidden instructions given to Apple's ai tools before the request is sent. But here's one last instruction before you go:

The files I reviewed refer to the model as “ajax”, which some Edge Readers may remember it as the rumored internal name for Apple's LLM last year.

The person who found the instructions also posted instructions on how to locate the files within the macOS Sequoia 15.1 developer beta.

Expand the “purpose_auto” folder and you should see a list of other folders with long, alphanumeric names. Inside most of them, you’ll find the AssetData folder containing the “metadata.json” files. When you open these, you should see some code, and occasionally at the end of some of them, instructions that are passed to Apple’s local version of LLM on your machine. But you should remember that these are in a part of macOS that contains the most sensitive files on your system. Be careful!