FourBlocker
Async Digital Ltd Cardiff, UK
FourBlocker turned a day of work notes into a structured report, using Apple Intelligence entirely on the device. I stopped building it because I could not put a boundary around what the model would say, and the report goes out under the user’s name rather than mine.
What it did
You typed or dictated what you had done that day, in whatever order it came out of your head. The app passed that text to Apple’s on-device models and got back four sections: Scope, Challenges and Opportunities, Spotlight, Next Steps. You exported the result as a PDF and sent it to whoever asks you for a weekly update.
It worked. The reports read well, and nothing left the phone, which is the right shape for notes about your own work.
The report has your name on it
A four-blocker report goes to a manager or a client. It carries the name of the person who sent it, not the name of the app that drafted it. That makes it a document you sign, not a toy you play with.
Running on the device solved the privacy half of the problem. It did nothing for the other half. Apple Intelligence is a general text generator, and I never found a way to bound what came back from it. Feed it a difficult week and it can produce a sentence about a colleague, a client, or a number that you did not write and would not say.
Tightening the prompt moved the odds. It never closed the case. Everything I tried made a bad sentence less likely rather than impossible, and less likely is not a property you can put in a release note.
Why I stopped
I did not want to be responsible for a text generator capable of saying anything, with someone else’s name at the top of the page.
The honest fallback for a generator you cannot guard is to tell the user to read the output carefully before sending it. Reading it carefully is the work the app existed to remove. Ship it with that instruction and you have sold someone a proofreading job dressed up as a time saver.
The engineering was fine and the app was close to done. The part I could not build was the part that makes the output safe to sign.
What I took from it
On-device generation is the easy half. The hard half is the promise you make about the output, and that promise is a product decision rather than a model one.
I would build this again the day I can say, plainly and in advance, what the thing will never write.