Prompting Apple Intelligence and Siri

TL;DR: As of September 3, 2026, the Siri overhaul Apple demoed (personal context, cross-app actions, a dedicated Siri app) is still coming later this year, in beta, English-only, and not yet in the EU. What's live today is the current Siri plus an opt-in ChatGPT extension. Three different systems handle your request: on-device processing, Apple's own Private Cloud Compute, and, only if you turn it on, OpenAI's ChatGPT. Prompting Siri well means working with voice-first, single-request input, not writing a longer ChatGPT prompt and hoping it translates.

What has Apple Intelligence actually shipped, and what's still coming?

Apple's own Apple Intelligence and Siri page, fetched September 3, 2026, leads with this exact framing: "New Apple Intelligence features coming this fall. Siri AI coming in English later this year." The footnote is more specific: the reimagined Siri "will be available In beta later this year and requires an Apple Intelligence–enabled device set to a supported language. Available in English to start." It also states plainly that Siri AI "will not initially be available in the EU on iOS, iPadOS and watchOS."

That matters because most of what circulated after the keynote preview described a Siri that understands personal context, takes actions across Messages, Music and Reminders based on what you're doing in the moment, and lives in its own dedicated app you can return to on any device. None of that is generally available today. It's a real, specific set of features Apple has committed to shipping; it just hasn't shipped yet, and it's gated by beta status, one language, and a region carve-out.

What's actually live right now, per the same page: Image Wand in Notes and Live Translation across Messages, FaceTime, Phone and AirPods are both labeled available now. Proofread-as-you-type, Call Context, and a description-driven calendar-event editor are each labeled coming in English, a step ahead of the beta-gated Siri rebuild but still not universally shipped. Genmoji, Image Playground, and the current generation of Visual Intelligence are already part of the existing release.

It's worth separating two different language lists here, because conflating them is an easy mistake. The existing generation of Apple Intelligence, the one already shipping, supports sixteen languages including English, French, German, Japanese, Korean, and both Chinese scripts, per Apple's own compatibility footnote. The rebuilt Siri that's still in beta is narrower: English only, to start, with the broader language rollout unannounced. If you're prompting in another supported language today, that already works for the current feature set; it just won't carry over automatically to the new Siri the day it ships.

What are the three different layers, and why does the difference matter for what you say to Siri?

This is the distinction that gets flattened in almost every generic how to prompt Apple Intelligence post, and it's the actual hazard: on-device processing, Private Cloud Compute, and ChatGPT are three different systems with three different privacy properties, and only one of them is a third-party AI model.

On-device processing is what Apple leans on first: your personal information is used without being collected, because the model runs locally. For requests too complex for the device, Apple's own page puts it this way: "And with groundbreaking Private Cloud Compute, Apple Intelligence can draw on larger server-based models, running on Apple silicon, to handle more complex requests for you while protecting your privacy." Private Cloud Compute itself is documented in more depth on Apple's own security research site. Apple's security team states the underlying guarantee more strongly in its own engineering write-up: PCC is built so that personal user data sent to it "isn’t accessible to anyone other than the user — not even to Apple."

ChatGPT is a different thing entirely, and it's opt-in on top of the other two. You turn it on yourself, in Settings, under Apple Intelligence & Siri. When it's on, Siri can hand a request to ChatGPT (Apple's own support guide gives the example "for example, insights about photos and documents"), and Writing Tools, Image Playground, visual intelligence with Camera Control, and Shortcuts can all call it too. Apple's documentation is specific that you're always asked to confirm before any photo or file is sent to ChatGPT, that your IP address is obscured, and that OpenAI does not receive anything tied to your Apple Account. You can use the extension without a ChatGPT account at all; signing in only changes whether your requests join your ChatGPT chat history.

Context window management, in the sense a ChatGPT or Claude user thinks about it, doesn't really apply to any of these three. None of them show you a scrollable conversation buffer you're expected to manage — that's the next thing to get right.

Why is prompting Siri different from prompting a chat model?

Four constraints, and they're structural, not stylistic:

1. Voice-first. Most Siri requests are spoken, not typed. A prompt built for a chat window (multiple sentences, a numbered list of constraints, a pasted document) doesn't have an equivalent shape when you're saying it out loud. Apple does support typing to Siri, but the interaction is still one short request, not a drafted message.
2. No pasted context. There's no field to paste a paragraph of background before your actual question, the way you'd paste a job description before asking ChatGPT to tailor a resume bullet. If Siri needs context, it has to already be on your device (a document open in Files, a photo open in Photos), or it has to ask you directly.
3. No visible context window. A chat app shows you the whole conversation so you can see what the model still remembers. Siri doesn't expose that. The new Siri app promises a place to return to past conversations, but you still can't see, at request time, how much of an earlier exchange is actually feeding the current answer.
4. Routing, not one model. A request to Siri can be handled entirely on-device, escalated to Private Cloud Compute, handed to ChatGPT, or routed into a specific subsystem: visual intelligence for anything involving the camera, Writing Tools for composing text, Shortcuts for automations. You don't choose the route; the phrasing and context of your request do.

That last point is the one worth building a habit around: name the target explicitly when you want a specific subsystem. "Hey Siri, ask ChatGPT to compose a haiku about dragons" (Apple's own example) routes straight to the ChatGPT extension instead of leaving Siri to decide whether ChatGPT would help. The same discipline (say what you want acted on, and where) is the core idea behind structuring any prompt at all; see what prompt engineering actually is if that's new ground.

How do you actually get better answers out of Siri today?

• One request, one intent. Siri parses a single spoken or typed request at a time; a request that bundles three separate asks ("remind me to call the dentist and also what's the weather and also text mom I'm running late") is likelier to catch only the last piece than to execute all three cleanly.
• Name the app or extension you want. "Ask ChatGPT to…" for the ChatGPT extension specifically. "Hey Siri, remind me to…" when you want the Reminders action, not a ChatGPT-style answer. Vague phrasing forces Siri to guess the route.
• Use Type to Siri for anything precise. Names, numbers, and exact phrasing (a person's name, a product SKU, an exact quote) are more reliable typed than spoken, since there's no autocorrect-style repair loop the way there is in a text field.
• Let visual intelligence carry the context instead of describing it. "What kind of recipes can I make with this?" pointed at a photo of fruit (Apple's own example) does more work than describing the ingredients verbally, because the image itself is the context.
• Confirm or deny deliberately. When Siri asks whether to hand a request to ChatGPT, that's a real decision point, not a formality: it's the moment you're choosing between Apple's infrastructure and OpenAI's for that specific request.

What should you not expect Siri to do yet?

Anything that depends on the personalized, cross-app Siri: reading your own photo library or email inbox by description, taking an action inside Messages or Music based on what you're doing right now, or picking up a conversation across devices in a dedicated Siri app. Apple's own page still lists these under the feature set that is coming later this year, in beta. As of September 3, 2026 it has not shipped anywhere, and it will be English-only once it does. If you want that kind of continuity today, the honest answer is that no combination of prompting gets you there yet; it's a shipped-features gap, not a prompting one. For a chat-model equivalent of "how much should I even ask for in one go," when to start a new conversation versus keep going covers the version of this question that does apply to tools with a real context window.

Copy-paste starting points

For the ChatGPT extension, routed explicitly:

Hey Siri, ask ChatGPT to [task], and keep it to [length/format constraint].

For visual intelligence, pointed at what's on your screen or in front of the camera:

[Point camera or take a screenshot] → "What is this, and [specific question]?"

For a Shortcut you want built by description rather than assembled manually:

Describe a shortcut that [trigger] and then [action], every time [condition].

Each of these works with the constraints above rather than against them: one intent, a named target, and context supplied by what's already on the device instead of pasted in. If you'd rather draft the fuller version of a request first (the way you would for ChatGPT or Claude) and then compress it down to one spoken sentence for Siri, that compression step is exactly what a prompt enhancer is for: write the detailed version once, then cut it down for voice.

The short version

The rebuilt Siri that got the keynote attention is still in beta, English-only, and missing from the EU as of this writing. Check apple.com yourself before you plan a workflow around it, since Apple has moved this date before. What's shipped is the current Siri plus an opt-in ChatGPT extension layered on top of two genuinely different privacy models, on-device and Private Cloud Compute. Prompting either well means working with voice-first, single-intent, no-pasted-context constraints, not writing a chat-model prompt and hoping it survives translation into speech.