Claude Code: turn a Swift finding into a tested patch

Keep the first task deliberately small

When an agent reviews an entire app at once, it can be difficult to tell which suggestions are worth acting on. Start with the Swift file you just changed. Define the behavior it must preserve, then ask for one focused review before allowing edits.

Connect the local tools

claude mcp add ios-agent -- npx -y ios-agent-mcp@latest

Run the command in your project context, reconnect Claude Code and confirm the server is available. The local server needs Node.js 20 or later. Xcode operations additionally need a Mac with Xcode. Claude Desktop has a different configuration path; use the client setup guide.

Ask for evidence before a rewrite

Review the concurrency in the Swift file I changed.
Show each finding with its location and supporting context.
Propose one minimal patch. Preserve the public behavior.
After approval, build and run the relevant tests.
Report failures and checks you could not run.

Read the suggested diff. A heuristic finding about actor isolation is a reason to inspect the surrounding code, not a reason to annotate every type. Keep unrelated refactors out of the patch so a failing test has a smaller set of possible causes.

Stop at a meaningful checkpoint

Save the changed files, commands used and remaining failures. If the build cannot run because Xcode or the scheme is unavailable, record that constraint instead of claiming success. The repository's Claude record demonstrates a bounded text-repair test; it does not establish complete app generation.

Official Claude MCP documentation · Project verification record

Decision checkpoints