Website illustration, followed by a real simulator screenshot of the Reading List demo.
I maintain iOS Agent Skill, an open-source collection of Swift guidance, local references and MCP tools. This post was prepared with AI assistance. Here is a small verification exercise you can use with or without the toolkit.
Use a reading list as the test case
A list that renders is only the starting point. Give your coding agent five acceptance checks:
- Add a book, terminate the app, and relaunch. The book should still exist.
- Mark a book finished and repeat the relaunch check.
- Search for an existing title, then a title that is absent. Both states should be understandable.
- Inject a failed save. The UI should surface the error rather than reporting success.
- Inspect empty and populated screens with accessibility in mind: meaningful labels, readable text and usable controls.
Use a temporary data store or synthetic fixture for failure cases. Do not test destructive behavior against your real library.
Retrieve guidance before editing
Search for the specific persistence or concurrency question. Read only the relevant guide section or source example. iOS Agent Skill includes local retrieval tools for this purpose; the repository is a dated reference snapshot, so verify changing APIs against official Apple documentation and your installed SDK.
Separate review findings from proof
The npm package ios-agent-mcp combines Swift review tools, Apple reference retrieval, scaffolding and simulator workflows. Its static reviews are heuristics. A finding deserves investigation; a clean review does not prove the app works.
Compile the project, run tests for the acceptance checks, and inspect simulator behavior. Record which checks actually ran and which are blocked by the environment. Local simulator operations require macOS and Xcode. Scaffolding creates starter files, not a finished app from a prompt.
Inspect a concrete example
The repository contains an editable Reading List demo with persistence, search, tests and screenshots. Use its README to reproduce the checks on your own Mac before extending it. Evidence for that example is not a guarantee about all generated apps.
Installation and client-specific setup are on the project website. Start with a small project and follow the setup for your coding client. The source is MIT licensed; no hosted simulator is included.
I have not established benchmarked token savings or guaranteed autonomous app completion. What I want to improve is the inspectable workflow between an idea, a code change and a verified result.
If you try it, a missing guide or reproducible false-positive review is useful feedback. Please keep private app code, credentials and raw logs out of public issues. Which of these five checks would you add to your current workflow first?
Top comments (0)