Intel AI PC Vision Skill

面向 Intel AI PC 的本地图像分类 Skill。通过 OpenVINO Runtime 调度 Intel CPU、GPU 和 NPU，使用 MobileNetV3-Small FP16 完成 JPG、JPEG、PNG 图片推理，并向 WorkBuddy、Qoder、TRAE Work 等 AI Agent 返回实际执行设备、推理延迟、模型精度和 Top-5 分类结果。

Intel AI PC Vision

Run local image classification as an Agent Skill. Return only actual OpenVINO output; never invent hardware availability or benchmark numbers.

Hardware Capability

Designed for Intel AI PC.

Supported acceleration:

Intel CPU
Intel GPU
Intel NPU through OpenVINO

Runtime: OpenVINO
Model: MobileNetV3-Small
Precision: FP16

AUTO selects the first available device in this order: NPU, GPU, CPU. If the requested device is unavailable, fall back to a real available device and report the selected device in JSON.

Execute

Resolve every bundled path from the directory containing this SKILL.md. WorkBuddy exposes that directory as ${CODEBUDDY_SKILL_DIR}. For a user-provided image on Windows, run:

powershell -ExecutionPolicy Bypass -File "${CODEBUDDY_SKILL_DIR}\scripts\run_skill.ps1" -ImagePath "C:\absolute\path\to\image.jpg" -Device AUTO

For a safe verification using the bundled competition sample, run:

powershell -ExecutionPolicy Bypass -File "${CODEBUDDY_SKILL_DIR}\scripts\run_skill.ps1" -ImagePath "${CODEBUDDY_SKILL_DIR}\assets\examples\dog.jpg" -Device AUTO

Agent Response

Read the JSON from standard output. Present:

actual device
measured latency_ms
model_precision
Top-5 labels and confidence values

State that latency is a measurement from the current machine. Do not claim NPU execution unless the JSON says "device": "NPU".

Read references/usage.md when installation or dependency details are needed. Read references/agent-demo.md for the tested WorkBuddy prompt and references/test-report.md for competition evidence.

Expected JSON contract:

{
  "image": "dog.jpg",
  "device": "GPU",
  "latency_ms": 5.2,
  "model_precision": "FP16",
  "top5": [
    {"label": "Blenheim spaniel", "confidence": 0.86}
  ]
}