[Feature] Skill for Android developer workflow automation

[AndreSand]

Note

Feature requests are accepted only if the proposed skill fills a verified gap in state-of-the-art LLMs. To assess this, please test your use case against both a fast model (e.g., Gemini 3 Flash) and a reasoning/thinking model (e.g., Gemini 3.1 Pro).

What is the gap in the LLM's knowledge?

While modern LLMs excel at generating code from isolated text prompts, they operate in a vacuum—they lack situational awareness of the local runtime environment and cannot natively verify their own visual output. Specifically, the gap consists of:

• Disconnection from the Local Toolchain: The LLM cannot independently compile code, boot an emulator, deploy an application, or navigate to a specific UI state to see if its changes actually work.

  • Lack of Visual Feedback Loops: The model cannot perceive the visual layout of its code changes in real time. It relies entirely on the developer to manually review the UI, spot layout defects, and feed that feedback back into the prompt loop.

  • Design-to-Code Context Fragmentation: The LLM cannot natively correlate static textual requirements (PRDs) with spatial visual assets (Figma mockups) and local implementation files all at once without manual orchestration.

Proposed Skill

I am proposing the skill /verify-implementation, a custom skill that acts as an autonomous loop, tying code generation, local environmental execution, and visual verification into a single atomic action.

The skill covers three core operational phases:

• Multimodal Grounding: It simultaneously ingests product requirement text (PRD) and visual design references (Figma/images) to understand the exact feature scope and target UI layout.

• Toolchain Orchestration: After applying the local source code modifications, the skill takes control of the local host. It kicks off parallel background tasks to compile the project (via Gradle) and prepare the runtime environment (booting an Android emulator/device via ADB), Android CLI.

• Automated Visual Regression: It automatically launches the target application screen, takes a live screenshot of the runtime state, and uses its vision capabilities to compare it directly against the original Figma mockup—identifying layout bugs or visual drift before presenting the work to the human developer.

Additional Context

The manual development loop—writing code, waiting for long compilation cycles, launching emulators, manually navigating to the changed screen, and eyeballing padding/alignment—is one of the largest time-sinks in mobile engineering.

By shifting the model from a basic "code generator" to an environmental executor, /verify-implementation changes the paradigm from human-in-the-loop verification to human-in-the-loop review. By the time the developer looks at the task, the agent has already attempted the change, built the app, caught visual discrepancies, and self-corrected—leaving the engineer with a clean diff and an objective, automated visual PASS/FAIL report.

Note: I have already built a working proof of concept that I use in my current workflow. I am happy to share a quick demo to show it in action.