Machine Learning Claude Skills

Train and optimize AI agents using Microsoft's Agent Lightning framework with reinforcement learning. Use when setting up agent training, instrumenting agents with tracing, configuring LightningStore, implementing reward functions, or optimizing prompts with RL/APO algorithms.

Expert in shipping production-grade AI-powered features — LLM integration patterns, RAG architecture, prompt engineering that scales, AI UX that users trust, safety and guardrails, streaming, and cost optimization.

Knowledge graph memory backend powered by Cognee. Query, status check, dataset management, and backend switching between Brain and Cognee. Triggers on: 'cognee', 'knowledge graph', 'graph memory', 'switch memory', 'memory backend'.

Read the recent M0 operational thread for a project, newest first, to pick up work started in this or another tool. Answers 'where were we' and 'what is next' before asking the user to repeat context. Local SQLite, no embeddings, no network. Triggers on: 'm0 recall', 'where were we', 'what did we do last time', 'catch me up', 'resume context', 'read the thread', 'what is next'.

Search the web for real-time information using the Skywork web search API. Use this skill whenever the user needs up-to-date information from the internet — for example, researching a topic, looking up recent events, finding facts or statistics, gathering material for a document or presentation, or answering questions that require current data. Also trigger when the user says things like "search for" / "搜索" / "検索" / "검색", "look up" / "查询" / "調べる" / "조회하다", "find information about" / "查找关于……的信息" / "……に関する情報を探す" / "…에 대한 정보를 찾다", "what's the latest on" / "……最新进展" / "……の最新情報" / "…의 최신 소식", or any request that implies needing information beyond your training data.

Comprehensive evaluation patterns for AI agents including multi-turn conversation testing, LLM-as-judge frameworks, benchmark suites, regression detection, and systematic eval pipelines for measuring agent quality and safety.

Test LLM, RAG, MCP, and agentic systems end to end. Build golden datasets, run deterministic checks and LLM judges, score retrieval, probe prompt injection, verify tool use, and gate CI on thresholds. Orchestrates DeepEval, Ragas, promptfoo, and Langfuse.

Instrument LLM apps with Langfuse tracing, then use traces, scores, and datasets to test in production, run evaluations on real traffic, catch regressions, and close the loop from incident to golden dataset.

Grade LLM and agent traces with OpenAI Evals - build datasets, configure string/python/model graders, run eval suites, and gate agent behavior changes in CI.

Evaluate and red-team LLM applications with promptfoo, declarative YAML evals, assertions, model comparisons, and automated adversarial scans for prompt injection, jailbreaks, PII leaks, and unsafe outputs in CI.

Measure RAG pipeline quality with context precision/recall, faithfulness, answer relevancy, and groundedness using Ragas and DeepEval, with golden datasets and pass/fail thresholds.

Gate RAG pipelines in CI with versioned golden eval sets, per-metric thresholds, baseline drift detection, and a build that fails when retrieval or answer quality regresses.

Evaluate RAG pipelines with Ragas, measuring faithfulness, answer relevancy, context precision and recall, building golden datasets, and wiring threshold gates into CI for retrieval regressions.

> Use when integrating LLMs, chat clients, embeddings, RAG pipelines, or AI agents into Spring Boot. Covers Spring AI ChatClient, prompt templates, embeddings, vector stores, and structured output. Use when user mentions Spring AI, LLM, ChatGPT, Claude, RAG, embeddings.

> Use when integrating LLMs, chat clients, embeddings, RAG pipelines, or AI agents into Spring Boot. Covers Spring AI 2.0 ChatClient, prompt templates, embeddings, vector stores, and structured output. Use when user mentions Spring AI, LLM, ChatGPT, Claude, RAG, embeddings.

Patterns for Bayesian inference in R using brms, including multilevel models, DAG validation, and marginal effects. Use when performing Bayesian analysis.

R performance best practices including profiling, benchmarking, vctrs, and optimization strategies. Use when optimizing R code.

Keeps ingestion repository and published GPT-RAG documentation aligned with shipped formats, configuration, indexing, deployment, audit, and operator behavior.

Prepares and validates gpt-rag-ingestion releases. Use for release branches, VERSION, changelog entries, compatible GPT-RAG pins, tags, images, and GitHub Release notes.

Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.

Profile a DeepStream pipeline with Nsight Systems and derive its configs from the measurement. Use when the user asks for an efficient, performant, or profiled pipeline — or to benchmark, tune, or measure FPS.

Produce a professional training/workshop report as a .docx file. Use this skill whenever the user mentions "training report", "workshop report", "compte rendu", "compte rendu de formation", "formation report", "debriefing a workshop", "write up a training session", "résumé de formation", or any request to document a training session, workshop, or onboarding event with individual participant feedback and recommendations. Also trigger when the user says things like "I just ran a workshop and need to write it up", "help me summarize what happened in my training session", or "I need to report back to management about a session I ran". Always use this skill — for short or long sessions, across any discipline (technical, soft skills, creative, compliance, onboarding, etc.) — whenever a structured written deliverable about a training event is needed.

Build production-ready LLM applications, advanced RAG systems, and intelligent agents. Implements vector search, multimodal AI, agent orchestration, and enterprise AI integrations. Use PROACTIVELY for LLM features, chatbots, AI agents, or AI-powered applications.

Semantic image-text matching with CLIP and alternatives. Use for image search, zero-shot classification, similarity matching. NOT for counting objects, fine-grained classification (celebrities, car models), spatial reasoning, or compositional queries. Activate on "CLIP", "embeddings", "image similarity", "semantic search", "zero-shot classification", "image-text matching".

Aesthetic assessment and remix partner with trained visual taste. Provides structured design critiques using a 6-dimension scoring system inspired by VisualQuality-R1 chain-of-thought reasoning.

Expert in drone systems, computer vision, and autonomous navigation. Specializes in flight control, SLAM, object detection, sensor fusion, and path planning. Activate on "drone", "UAV", "SLAM", "visual odometry", "PID control", "MAVLink", "Pixhawk", "path planning", "A*", "RRT", "EKF", "sensor fusion", "optical flow", "ByteTrack". NOT for domain-specific inspection tasks like fire detection, roof damage assessment, or thermal analysis (use drone-inspection-specialist), GPU shader optimization (use metal-shader-expert), or general image classification without drone context (use clip-aware-embeddings).

Advanced CV for infrastructure inspection including forest fire detection, wildfire precondition assessment, roof inspection, hail damage analysis, thermal imaging, and 3D Gaussian Splatting reconstruction. Expert in multi-modal detection, insurance risk modeling, and reinsurance data pipelines. Activate on "fire detection", "wildfire risk", "roof inspection", "hail damage", "thermal analysis", "Gaussian Splatting", "3DGS", "insurance inspection", "defensible space", "property assessment", "catastrophe modeling", "NDVI", "fuel load". NOT for general drone flight control, SLAM, path planning, or sensor fusion (use drone-cv-expert), GPU shader development (use metal-shader-expert), or generic object detection without inspection context (use clip-aware-embeddings).

Selects the optimal LLM model and provider for each task based on complexity, cost budget, and capability requirements. Routes cheap tasks to Haiku/GPT-4o-mini and complex tasks to Sonnet/Opus/o1. Use when deciding which model to call, optimizing LLM costs, or building multi-model agent systems. Activate on "which model", "model selection", "route to model", "LLM cost", "model routing", "cheap vs expensive model". NOT for prompt engineering (use prompt-engineer), model fine-tuning, or training custom models.

Build production LLM streaming UIs with Server-Sent Events, real-time token display, cancellation, error recovery. Handles OpenAI/Anthropic/Claude streaming APIs. Use for chatbots, AI assistants, real-time text generation. Activate on "LLM streaming", "SSE", "token stream", "chat UI", "real-time AI". NOT for batch processing, non-streaming APIs, or WebSocket bidirectional chat.

Coaches end-to-end ML system design interviews covering inference pipelines, recommendation systems, RAG, feature stores, and monitoring. Use for L6+ design rounds, ML architecture whiteboarding, system design practice, serving tradeoff analysis. Activate on "ML system design", "ML interview", "recommendation system design", "RAG architecture", "feature store design", "model serving". NOT for coding interviews, behavioral questions, ML theory quizzes, or paper implementations.

Expert photography composition critic grounded in graduate-level visual aesthetics education, computational aesthetics research (AVA, NIMA, LAION-Aesthetics, VisualQuality-R1), and professional image analysis with custom tooling. Use for image quality assessment, composition analysis, aesthetic scoring, photo critique. Activate on "photo critique", "composition analysis", "image aesthetics", "NIMA", "AVA dataset", "visual quality". NOT for photo editing/retouching (use native-app-designer), generating images (use Stability AI directly), or basic image processing (use clip-aware-embeddings).

Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.

Expert in 3D computer vision labeling tools, workflows, and AI-assisted annotation for LiDAR, point clouds, and sensor fusion. Covers SAM4D/Point-SAM, human-in-the-loop architectures, and vertical-specific training strategies. Activate on '3D labeling', 'point cloud annotation', 'LiDAR labeling', 'SAM 3D', 'SAM4D', 'sensor fusion annotation', '3D bounding box', 'semantic segmentation point cloud'. NOT for 2D image labeling (use clip-aware-embeddings), general ML training (use ml-engineer), video annotation without 3D (use computer-vision-pipeline), or VLM prompt engineering (use prompt-engineer).

為使用者產生高品質的 AI 生圖、生影片、生音樂提示詞，並在需要時透過瀏覽器自動化實際送到目標平台。涵蓋 OiiOii、Kling 3.0/O-series、Seedance 2.0 pro、Suno v5.5、Seedream 5.0/4.0、Vidu Q3、Midjourney V8.1、Flux 1.1 Pro / Kontext、Runway Gen-4.5 / Aleph、Google Veo 3.1、Ideogram 3、Nano Banana Pro、Stable Diffusion 3.5（⚠️ OpenAI Sora 2 已於 2026-04-26 停運，API 撐到 2026-09-24，預設改推 Runway/Veo/Kling）。只要使用者提到「AI 生圖」「AI 影片」「AI 音樂」「做 MV」「做 storyboard」「寫 prompt 給 XXX」「我想用 Kling/Suno/Midjourney/Runway/Veo...」「幫我操作 OiiOii / 即夢 / 可靈」「txt2img / img2video / 文生圖 / 文生影片 / 圖生影片」「角色一致性」「多鏡頭分鏡」「運鏡」「結果有瑕疵 / 不夠精緻 / 怎麼修」，或任何跟上述平台或影像/影片/音樂生成工作流相關的任務，都要用這個 skill。即使他們沒講明平台，只要任務是要餵給某個生成模型的 prompt，就用這個 skill 幫他們選對的平台、寫對的格式。

>- MANDATORY pre-flight check before answering any question or writing any code related to Meta Quest VR headsets (Quest 2, Quest 3, Quest 3S, Quest Pro) or apps that target them. Forces verification against authoritative Meta sources via the metavr CLI / metavr MCP tools BEFORE relying on training-data knowledge. Counters the failure mode where agents answer Quest-related questions from stale memory and ship deprecated APIs, broken Android manifests, and store-rejected builds. Loads automatically when the user is in a Quest project (any reference to Oculus / Meta Quest / Horizon OS, Unity OVR or Meta XR packages, com.meta.* / com.oculus.* package IDs, Quest-targeted AndroidManifest, .meta files, or developers.meta.com / developer.oculus.com URLs).

Create personalized triathlon, marathon, and ultra-endurance training plans. Use when athletes ask for training plans, workout schedules, race preparation, or coaching advice. Can sync with Strava to analyze training history, or work from manually provided fitness data. Generates periodized plans with sport-specific workouts, zones, and race-day strategies.

Web search using the Brave Search CLI (`bx`). Use for ALL web search requests — including "search for", "look up", "find", "what is", "how do I", "google this", and any request needing current or external information. Prefer this over the built-in web_search tool whenever bx is available. Also use for: documentation lookup, troubleshooting research, RAG grounding, news, images, videos, local places, and AI-synthesized answers.

USE FOR RAG/LLM grounding. Returns pre-extracted web content (text, tables, code) optimized for LLMs. GET + POST. Adjust max_tokens/count based on complexity. Supports Goggles, local/POI. For AI answers use answers. Recommended for anyone building AI/agentic applications.

Decompose portfolio returns into explainable components to identify where value was added or lost. Use when the user asks about Brinson attribution, allocation vs selection effects, factor-based attribution, fixed-income attribution, or currency attribution. Also trigger when users mention 'what drove my returns', 'was it stock picking or sector bets', 'alpha decomposition', 'multi-period linking', 'interaction effect', 'active return breakdown', or ask why their portfolio outperformed or underperformed the benchmark.

> Deploy AI models to embedded hardware using MathWorks tools (MATLAB, Simulink, models rebuilt as dlnetwork for lean hardware (Cortex-M, DSP), (2) direct C/C++ code generation from PyTorch and LiteRT models for high-performance hardware (Cortex-A, x86, GPU). neural networks; compress AI models for MCU/DSP; integrate AI in Simulink for system-level simulation; import PyTorch/ONNX/TensorFlow models for embedded deployment; optimize AI for resource-constrained hardware; or use loadPyTorchExportedProgram, importNetworkFromPyTorch, dlquantizer, exportNetworkToSimulink, or Embedded Coder with AI models.

Develop, debug, and optimize SGLang LLM serving engine. Use when the user mentions SGLang, sglang, srt, sgl-kernel, LLM serving, model inference, KV cache, attention backend, FlashInfer backend, MLA, MoE routing, MoE dispatch, expert parallelism SGLang, speculative decoding, disaggregated serving, TP/PP/EP, radix cache, continuous batching, chunked prefill, CUDA graph SGLang, model loading, quantization FP8/GPTQ/AWQ, JIT kernel, triton kernel SGLang, DeepSeek serving, EPLB (expert load balancing), HiCache, launch_server, sglang Engine API, LoRA inference, torch.compile SGLang, or asks about serving LLMs with SGLang. Also use when the user wants to add a new model to SGLang, add a new attention backend, debug SGLang serving issues, or optimize SGLang throughput/latency.

> ESM protein language models for embeddings, sequence scoring, structure pseudo-log-likelihood (PLL) or mutation-effect scores, (2) Getting protein embeddings for clustering or filtering, (3) Predicting complex structures with ESMFold2, (4) Designing binders by inverting ESMFold2, (5) Filtering designs by sequence plausibility. For diffusion-based structure prediction, use boltz or chai. For QC thresholds, use protein-qc. For gradient-based multi-objective design, use mosaic.

Exhaustively profile a dataset and list ALL possible analyses — distributions, correlations, rankings, trends, group comparisons, anomalies. Reads detective.json for context. Runs after the Detective/Scout, before the Editor. Outputs analyst.json with ana_xx IDs and chart-ready data_tables.

Generate text embeddings via OpenRouter using Qwen3-Embedding-8B.

Front stage for /data2story-pro when the reader has no dataset — only a vague idea. Converges the idea into a concrete, data-backed topic through a sparring-partner dialogue (anti-sycophantic, feasibility-pressure-tested), then acquires a REAL dataset through find-data, with a user checkpoint after each. Returns a validated DATA_DIR for the main pipeline. Not a newsroom role — runs upstream of Detective, before any dataset exists. Real data only; never a reason to synthesize data.

Research external context for a dataset — domain background, history, related studies, and why this data matters. Outputs detective.json (structured findings with det_xx IDs) before any analysis begins.

Data Journalist Agent (Data2Story) — orchestrator: turn a dataset into a blog. Runs detective → analyst → editor → designer → programmer → auditor → inspector in sequence. Creates a versioned project folder for each run.

Exhaustively profile a dataset and list ALL possible analyses — distributions, correlations, rankings, trends, group comparisons, anomalies. Reads detective.json for context. Outputs analyst.json with ana_xx IDs and chart-ready data_tables.