A little while ago we wrote about building our AI apps around the small model Apple makes available on compatible devices. Current versions may use Apple’s built-in runtime where available or a compatible downloaded model on other supported iPhones and iPads. Whichever runtime is active, the model alone is not enough. The harness is the engineering around it that narrows the task, grounds the input, validates outputs, and leaves the user in control.
That’s easy to say. Here’s what the harness is actually made of: three tricks we lean on in every app.
1. Give it a form to fill, not a blank page
Ask a small model an open question and you get back a paragraph you then have to untangle, sometimes with the answer buried in polite chatter, sometimes in a shape your app can’t read.
So we do not ask for an unconstrained blob when a structured result will do. We hand the model a form with labeled fields and validate what comes back. In Better Writing, Rewrite mode returns the improved text and structured change information for the word-level diff; it is also built to reject unsupported details. The newer Draft and Reply workflows are broader, but they remain grounded in notes the user supplies or a screenshot transcript the user confirms. Structure reduces parsing failures and constrains the job. It does not make AI infallible, so the app still shows its work and asks the user to review important text.
2. Let it look things up, not remember
A small model’s memory is fuzzy. Ask it for a grammar rule or a travel phrase and it will half-remember, confidently.
So we don’t ask it to remember; we let it look up. The real answer lives in the app, and the model’s job shrinks to finding the right entry and using it. Think open-book versus closed-book exam: same student, far better answers with the book open. In Second Brain, the “book” is your own saved notes. That’s why its answers come with sources and can say “I don’t have that.” In Pocket Lingo, it’s a curated phrasebook instead of a phrase the model invented on the spot. The weak model plus the right facts beats the weak model guessing, every time.
3. Test every change before it ships
Here’s the quiet danger with AI: you tweak how you ask the model one small thing, it looks better on the example in front of you, and it silently gets worse on ten cases you didn’t check.
So each app keeps a set of real examples with known-good answers. Every time we change how the model is prompted, we run the whole set and compare the score. If the score drops, the change doesn’t ship. It’s a spell-check for our own work: it catches the “harmless little tweak” that would have quietly broken things, before it ever reaches your phone.
The through-line
None of these makes the model smarter. They make it safer to trust by giving it less room to go wrong. Constrain the shape of the answer. Hand it the facts instead of asking it to recall them. Prove it still works after every change.
Smaller and careful beats bigger and loose, especially when the subject is your writing, your memory, and your day.
Studio AM