Synthetic users
A brief. From the owner's voice memo of 13 November 2026, written up by the agent.
The problem
The app is built by an agent that cannot see it. Continuous integration compiles and tests it; the owner sees it on a device. Between the two there is nobody checking whether a person arriving for the first time can find the record button, understand what a vault is, or get back from the place a tap took them.
The idea
Invented users, each with a purpose and a level of knowledge, who walk the screens and report where they hesitated, guessed wrong or got lost. The newsroom at sgit.ai does this for a website, one screenshot at a time. The same applies here.
For example:
- someone who wants to record a memo while driving and has never heard of a vault;
- someone careful about privacy who wants to know exactly what leaves the phone;
- a developer who wants to see every request the app makes;
- someone who lost a recording in another app and does not trust this one.
How it would work
- Pictures of every screen. A workflow in continuous integration runs the app in a simulator, one launch per screen with sample data, and keeps a screenshot of each on an iPhone and an iPad. The sibling repository has this already; it is to be carried over. It also gives the agent its first sight of its own work.
- A walk. A synthetic user is given a task and the screenshots, and says at each step what it would tap and what it expects. Where its expectation and the next screen disagree is a finding.
- A record. Findings are kept as files, each with the screen, the user and what went wrong, and a status as review findings have elsewhere in this family.
Limits
A synthetic user reads a picture. It does not feel a button that is too small for a thumb, hear nothing, or drive a car. Simulators have no speech model and no Apple Intelligence, so screens that depend on them are shown with sample data. The owner on a device remains the test that counts.