iOS Simulator Claude Skill (github.com)

Conor Luddy released an experiment — an “ios-simulator-skill” for Claude Code that enables accessibility‑first, semantic automation of iOS Simulator workflows. The package is a lightweight skill (12 production scripts) that automates builds, navigation, gestures, text entry, accessibility audits, visual diffs, real‑time log monitoring and full app‑state capture. It’s zero‑config on macOS with Xcode, Python 3, and optional idb, and Claude Code automatically loads the skill when your request implies simulator work. Outputs are token‑efficient by design (3–5 lines by default, ~97% token savings versus raw tools) with progressive disclosure via --verbose and --json flags for deeper debugging. Significance for AI/ML practitioners: this turns brittle pixel taps into robust, semantic actions that are easier for agents to plan, verify and troubleshoot, reducing token costs and noise when integrating agentic testing into CI, dataset collection or automated UX/regression pipelines. Key technical details: navigator uses fuzzy/text/type matching; screen_mapper returns concise UI trees; accessibility_audit reports WCAG issues as structured JSON; visual_diff supports pixel thresholds; app_state_capture bundles screenshot, UI tree and logs for reproducible bug reports. The author notes it’s a spike (they still favor XC‑MCP for predictability) but positions the skill as a practical, agent‑friendly bridge between LLMs and reliable iOS automation.