Today we submitted three apps to the App Store: Better Writing, Second Brain, and Pocket Lingo. Their AI processing runs on-device rather than on a Studio AM server, and none requires a Studio AM account. Some storage options have evolved since launch: for example, Second Brain can optionally sync supported note data through the user’s private iCloud while keeping AI processing on-device.

The original launch path centered on the small language model Apple ships on Apple Intelligence devices. Current versions can also offer a downloaded on-device model on other supported hardware. Here is the honest part: whichever small model is available, the model on its own is not very smart.

The bet

When we first tried Apple’s on-device model, it failed in almost every way a model can fail. We asked it to fix the grammar of “how gas cost” and it answered the question instead of fixing the sentence. It invented facts. It told us Bonjour is pronounced “bohn-ZHOO-roh.” Asked to make a quiz, it offered the right answer as all four options.

The obvious conclusion is “wait for a bigger model.” We made a different bet: a weak model plus a strong harness beats a big model with none.

A harness is everything around the model. The instructions that give it exactly one bounded job. The code that checks every output before a person sees it. The curated content that never came from a model at all. The evaluation suite that catches a regression before it ships.

What the harness looks like

A few real examples from this week:

Guard against unsupported details. Better Writing’s Rewrite mode is built to preserve meaning and reject unsupported additions. It shows the word-level changes so the user can inspect the result; as with any AI suggestion, important text still needs review. In Ask Monie’s new Magic Log, an amount is accepted only if the same digits appear in what you typed. A model cannot make up “$12.50” for your lunch.

Honesty over confidence. Second Brain answers questions from your own notes, and every answer shows its sources. When it doesn’t know, it says “I don’t have that in your notes yet.” We measured the alternative (a confident wrong answer about your own memory) and decided it’s the one thing this kind of app must never do.

Don’t ask the model for what it’s bad at. Pocket Lingo never asks the model how to pronounce anything, because its pronunciation guides are fiction. Instead, the phrase speaks aloud with each word highlighted as it’s read. The phrasebook itself is written and checked by people, not generated on your phone.

Test on the real thing. Every one of these behaviors was verified on an actual iPhone before shipping, with small evaluation suites we run every time a prompt changes. When the model got “who is April?” embarrassingly wrong, the fix wasn’t a better vibe; it was a measured 10-out-of-12 becoming 12-out-of-12.

Why on-device is the point

We could have wired these apps to a big cloud model and gotten smarter answers on day one. We didn’t, for one reason: the whole product promise is that your words stay yours. What you write, what you remember, what you spend, what you practice saying: that’s about as personal as data gets. For us, on-device is the feature, not a technical footnote.

The trade is real, but the reach is wider than the Apple Intelligence device list alone. Current versions use Apple’s built-in on-device model where it is available, and some can offer a compatible downloaded model on other supported iPhones or iPads. The exact model size and memory floor differ by app, so the current product page and in-app setup, not a rough phone-generation promise, are the source of truth. After compatible setup, core AI work can run offline; downloads, purchases, optional sync, and explicit external actions can still need a connection.

Better Writing, for example, requires iOS 26 or iPadOS 26. Its lighter downloaded model is about 0.9 GB and needs at least 3.5 GB of physical memory; its enhanced model is about 2.6 GB and needs at least 5.5 GB. Devices below that downloaded-model floor need Apple’s compatible built-in runtime.

Where this goes

The three new apps taught us a repeatable craft, and it’s already spreading to the apps we had: Ask Monie’s Magic Log now turns “coffee 4.50 at Blue Bottle” into a filled-out entry you just confirm. The maintenance-log apps are next.

Small studio, small model, strong harness. It turns out that’s enough.

Studio AM