What problem does it solve? AI agents normally cannot see whether their UI edits actually worked on a running Android device. This Skill gives an MCP-capable agent (Claude Code, Cursor) a closed feedback loop over Compose HotSwan: edit a Kotlin file, trigger a hot reload, capture a device screenshot, and evaluate the result against the design intent without a human steering each step. ## Core Features & Use Cases - Seven verbatim MCP tools: hotswan_get_status, hotswan_reload, hotswan_take_screenshot, hotswan_start_snapshot, hotswan_stop_snapshot, hotswan_select_variant, and hotswan_build_and_install, with rules against inventing tool names. - Canonical agent loop: status check, snapshot session, edit, reload with tier reading, screenshot evaluation, and optional rollback to a preferred variant. - Safety habits: watcher-state verification before edits, snapshot wrapping for rollback, and fallback to full install for schema changes like new parameters or resource IDs. - Use Case: A developer tells Claude Code to tune a Compose screen until it matches a mock; the agent iterates autonomously, rolls back to the best snapshot variant, and only falls back to a full install when an edit changes the class schema. ## Quick Start Ask your MCP-connected AI agent to check the HotSwan watcher status, start a snapshot session, then iteratively edit the target Kotlin composable, reload, and screenshot until the screen matches your design mock.
