On-Device Apple ML: The Core-ML Skill for Agents

Where this fits

A developer is building a photo app that must tag objects on-device, or a text app that needs sentiment analysis without sending user data to the cloud. The core-ml skill activates in that moment: it advises whether to use a Core ML model, Create ML for training, the Vision framework for images, or NaturalLanguage for text — and it covers the "Core ML vs Foundation Models" decision before you write any integration code. It turns the agent from a generic Swift assistant into one that understands Apple's ML ecosystem.

Why agents benefit

• Makes the Core ML vs Foundation Models decision explicit, so the agent doesn't default to a cloud API when on-device is the right call.
• Knows the Xcode integration path: adding .mlmodel files, generating model classes, and wiring them into Swift code.
• Can use Vision and NaturalLanguage correctly — object detection, face/body pose, text recognition, sentiment, language detection — without hallucinating framework APIs.
• Can generate Create ML training flows for custom models, including dataset preparation and evaluation.
• Covers optimization options like quantization and compression to keep app size and memory reasonable.

Practical scenarios

On-device object detection for a privacy-focused camera app

The user wants to detect pets in photos and never upload images. The skill recommends a Core ML object detection model plus the Vision framework for requests, and guides the agent in writing the Swift integration. Result: the agent produces working code and explains the model file size tradeoffs.

Sentiment analysis on user-submitted reviews

A journaling app wants to detect sentiment locally. The skill points to NaturalLanguage's sentiment analyzer instead of a heavy model, and provides the code to classify text with zero network calls.

Deciding between Core ML and a foundation model API

A developer wants to summarize documents. The skill lays out the tradeoffs: latency, privacy, model size, and update cadence. When a foundation model is required, the agent knows to stop and recommend a streaming API instead of forcing a local model.

Add it to your agent workflow

Install with:

npx skills add rshankras/claude-code-apple-skills --skill core-ml

Then prompt your agent with a concrete task, e.g., "Add on-device dog detection to my camera app using Core ML and Vision. Recommend a model and show the Swift code." The skill activates its advisory and generator workflows, producing a model recommendation and the integration code.

Compare before adopting

Before wiring core-ml into a regular workflow, compare it against alternatives like an LLM with broad Swift knowledge. Check quality signals: the skill is tagged as advisory/generator, it includes a review-by date (2027-06-22) which helps you judge freshness, and the GitHub repo shows genuine community traction at 696 stars. Verify it actually reduces iteration time for your specific Apple ML tasks rather than just suggesting the obvious.

Why it is worth tracking

Core-ml is a good signal of a growing category: skills that encode platform-specific expertise — Apple's ML stack, not generic coding advice. The maintainer reviews and version-stamps the skill, and the star count indicates real-world use. Track it if you regularly build iOS/macOS features involving vision, text analysis, or custom models, and evaluate it whenever your agent starts guessing about Core ML APIs. For a quick reference on the skill's intent and maintenance status, the OpenAgentSkill listing is a useful comparison point: https://www.openagentskill.com/skills/rshankras-core-ml