Machine Learning Skills

Same-epoch comparison of training runs across wandb, neptune, tensorboard, or mlflow. Aligns runs at the student's current step (never current-vs-final-of-baseline) and separates proxy metrics from downstream targets. Use when the user asks to compare runs, check if a run is improving, track lag against a baseline, rank experiments, or evaluate run-vs-run performance.

Pre-flight checklist for long-running ML training jobs covering config diff, run naming, path verification, monitoring setup, and restart-cleanup. Use when the user asks to launch, kick off, start, restart, or kill a training run, or mentions launching a multi-hour or multi-day GPU job (python train, accelerate launch, torchrun, deepspeed, sbatch, tmux training).

> Use when the user wants to analyze dataset bias, create stratified samples, evaluate fairness, or plan dataset collection. Triggers on phrases like "dataset bias", "stratified sample", "class imbalance", "data distribution", "fairness analysis", or "ethical review".

End-to-end paper reproduction from arxiv URL through smoke runs to replication experiments. Handles missing or partial official code, missing training scripts, missing hyperparameters, and private datasets via similar-public-dataset substitution. Use when the user asks to reproduce, implement, replicate, or re-run a paper from scratch, or pastes an arxiv URL with reproduction intent.

> Designs AI-powered API features, LLM tool/function definitions, MCP server tool schemas, natural language to API conversion, and agentic API workflows. Use whenever the user asks about "AI calling my API", "function calling schema", "tool definition for LLM", "MCP tools", "natural language API", "AI agent", "let Claude use my API", "OpenAI function calling", "Anthropic tool use", "API agent workflow", "LLM plugin", "AI integration", "RAG with my API", or "chatbot that calls my API".

>- Residual re-identification RISK CHECK on text you have ALREADY redacted (defensive, dual-use). Use when the user asks to "check residual re-id risk", "red-team my redaction", "what can an attacker still infer", "is this safe to share", or assess "re-identification risk" after anonymizing. Re-runs the CONFIDE detectors on the redacted output to surface surviving identifiers (singling-out), checks multiple files for linkability, and optionally probes a local model for still-inferable attribute CATEGORIES (inference) — mapped to GDPR Art-29. Reports risk categories/counts only,

Process textual and multimedia files with various LLM providers using the llm CLI. Supports both non-interactive and interactive modes with model selection, config persistence, and file input handling.

Run candidate product, brand, company, or benchmark names through an audition — authoritative domain-availability checks, collision research across SaaS/GitHub/packages/the target adjacent domain, a light trademark and ownability read, a ranked callback list, and an interactive casting report of finalists with optional draft branding. Use when the user is naming a product, app, company, feature, or benchmark; asks "is this name taken", "check these domains", "help me pick a name", "is X available", "name my product", "brand name research", "audition names"; or wants to compare and pressure-test a shortlist of candidate names before committing.

Iterate on RAG systems with structured evals instead of eyeballing. This skill should be used when the user is tuning a RAG pipeline — changing retrieval prompts, swapping models, adjusting chunking, or debugging poor answers — and wants a cheap, ranked set of experiments with cost tracking and structured feedback on the stack. Also use when the user asks "how do I know if my RAG is working?", "this RAG eval is burning money", or "what should I try next on retrieval?".

Index and search Claude Code sessions using semantic embeddings (Gemini). Find past sessions by topic, relaunch the best match. Triggers on "find session", "which session did I", "relaunch the session where", "session about X".

This skill should be used to watch a long-running background job (ffmpeg/media encode, qmd or other embedding/vector-DB run, batch agent/LLM pipeline, or a real-browser/agent-browser daemon) until it finishes or wedges, then deliver a verdict (done, needs-attention, or blocked) plus the exact next command, without burning dozens of manual poll commands. Triggers on "babysit this job", "watch this until it's done", "ping me when the encode/embed/batch finishes", "is this background process stuck", "monitor this ffmpeg/qmd run", or any request to wait on a long-running process and be told when it's complete or hung.

This skill should be used to generate realistic, persona-consistent synthetic coaching and therapy session transcripts for evals, demos, and training data. It produces fictional but believable coach/client (or therapist/client) dialogue grounded in a chosen modality (ICF/GROW coaching, CBT, IFS parts work, ACT/motivational interviewing) and exports to Fathom/Granola transcript style, plain dialogue, structured JSON, or Obsidian markdown. Triggers on requests like "generate a synthetic coaching session", "make fake therapy transcripts for evals", "create demo session transcripts", "synthetic CBT dialogue", "persona-consistent coaching transcript", "test data for my session summarizer", or "mock coaching call".

根据用户的视频需求(场景、时长、音频、语言、平台限制、预算、是否需要本地部署/角色一致性等),从 16 个主流 AI 视频模型(12 商业 + 4 开源)中推荐最匹配的 1-3 个,并解释为什么。覆盖商业:Seedance 2.0、HappyHorse 1.0、Kling 3.0、Sora 2、Veo 3.1、**Gemini Omni**(2026-05 新)、Runway Gen-4.5/Aleph、Pika 2.5、Hailuo 02、Hunyuan Video 1.5、Wan 2.7、即梦 AI;开源:LTX-Video 0.9.7、Mochi 1、CogVideoX 5B、Higgsfield Soul。用于"用哪个 AI 视频模型好"、"Sora 还是 Kling"、"国内有什么模型"、"哪个免费"、"哪个支持中文"、"哪个能编辑已有视频"、"哪个能本地部署"、"哪个角色一致性最强"等触发场景。

Use when attacking an AI/ML system or model — prompt injection & jailbreaks (Crescendo, Skeleton Key, Best-of-N), RAG/vector poisoning, agentic/MCP exploitation (CVE-2025-54136), ML supply-chain RCE (pickle CVE-2025-32434), model extraction / membership inference / adversarial suffixes (GCG)

Establishes CPU/GPU baselines before resource-intensive operations. Use before builds, training runs, or any task that pins cores or GPUs for over a minute.

Routes multi-tool workflows through MCP servers for large datasets and pipelines. Use when Bash tool overhead is limiting throughput on data-heavy tasks.

Provisions the oracle ML inference daemon with onnxruntime via uv. Use when setting up local ONNX model inference for skill quality evaluation.

|

Use this skill when working with scientific research tools and workflows across bioinformatics, cheminformatics, genomics, structural biology, proteomics, and drug discovery. This skill provides access to 600+ scientific tools including machine learning models, datasets, APIs, and analysis packages. Use when searching for scientific tools, executing computational biology workflows, composing multi-step research pipelines, accessing databases like OpenTargets/PubChem/UniProt/PDB/ChEMBL, performing tool discovery for research tasks, or integrating scientific computational resources into LLM workflows.

Retrieval-augmented generation (RAG) skill for the D&D 5e System Reference Document (SRD). Use when answering questions about D&D 5e core rules, spells, combat, equipment, conditions, monsters, and other SRD content. This skill provides agentic search-based access to the SRD split into page-range markdown files.

Implement ReasoningBank adaptive learning with AgentDB's 150x faster vector database. Includes trajectory tracking, verdict judgment, memory distillation, and pattern recognition. Use when building self-learning agents, optimizing decision-making, or implementing experience replay systems.

Achieve aggressive v3 performance targets: 2.49x-7.47x Flash Attention speedup, 150x-12,500x search improvements, 50-75% memory reduction. Comprehensive benchmarking and optimization suite.

Implements document chunking, embedding generation, vector storage, and retrieval pipelines for Retrieval-Augmented Generation systems. Use when building RAG applications, creating document Q&A systems, or integrating AI with knowledge bases.

Provides AWS CloudFormation patterns for Amazon Bedrock resources including agents, knowledge bases, data sources, guardrails, prompts, flows, and inference profiles. Use when creating Bedrock agents with action groups, implementing RAG with knowledge bases, configuring vector stores, setting up content moderation guardrails, managing prompts, orchestrating workflows with flows, and configuring inference profiles for model optimization.

Provides chunking strategies for RAG systems. Generates chunk size recommendations (256-1024 tokens), overlap percentages (10-20%), and semantic boundary detection methods. Validates semantic coherence and evaluates retrieval precision/recall metrics. Use when building retrieval-augmented generation systems, vector databases, or processing large documents.

Provides Amazon Bedrock patterns using AWS SDK for Java 2.x. Invokes foundation models (Claude, Llama, Titan), generates text and images, creates embeddings for RAG, streams real-time responses, and configures Spring Boot integration. Use when asking about Bedrock integration, Java SDK for AI models, AWS generative AI, Claude/Llama invocation, embeddings for RAG, or Spring Boot AI setup.

Provides patterns to build declarative AI Services with LangChain4j for LLM integration, chatbot development, AI agent implementation, and conversational AI in Java. Generates type-safe AI services using interface-based patterns, annotations, memory management, and tools integration. Use when creating AI-powered Java applications with minimal boilerplate, implementing conversational AI with memory, or building AI agents with function calling.

Provides Retrieval-Augmented Generation (RAG) implementation patterns with LangChain4j for Java. Generates document ingestion pipelines, embedding stores, vector search, and semantic search capabilities. Use when building chat-with-documents systems, document Q&A over PDFs or text files, AI assistants with knowledge bases, semantic search over document repositories, or knowledge-enhanced AI applications with source attribution.

Provides integration patterns for LangChain4j with Spring Boot. Configures AI model beans, sets up chat memory with Spring context, integrates RAG pipelines with Spring Data, and handles auto-configuration, dependency injection, and Spring ecosystem integration. Use when embedding LangChain4j into Spring Boot applications, building Java LLM applications with @Bean configuration, or setting up Spring AI patterns.

Provides configuration patterns for LangChain4J vector stores in RAG applications. Use when building semantic search, integrating vector databases (PostgreSQL/pgvector, Pinecone, MongoDB, Milvus, Neo4j), implementing embedding storage/retrieval, setting up hybrid search, or optimizing vector database performance for production AI applications.

Provides Qdrant vector database integration patterns with LangChain4j. Handles embedding storage, similarity search, and vector management for Java applications. Use when implementing vector-based retrieval for RAG systems, semantic search, or recommendation engines.

Enables interaction with Google NotebookLM for advanced RAG (Retrieval-Augmented Generation) capabilities via the notebooklm-mcp-cli tool. Use when querying project documentation stored in NotebookLM, managing research notebooks and sources, retrieving AI-synthesized information, generating audio podcasts or reports from notebooks, or performing contextual queries against curated knowledge bases. Triggers on "notebooklm", "nlm", "notebook query", "research notebook", "query documentation in notebooklm".

Stop and consult this skill whenever your response would include specific facts about Anthropic's products. Covers: Claude Code (how to install, Node.js requirements, platform/OS support, MCP server integration, configuration), Claude API (function calling/tool use, batch processing, SDK usage, rate limits, pricing, models, streaming), and Claude.ai (Pro vs Team vs Enterprise plans, feature limits). Trigger this even for coding tasks that use the Anthropic SDK, content creation mentioning Claude capabilities or pricing, or LLM provider comparisons. Any time you would otherwise rely on memory for Anthropic product details, verify here instead — your training data may be outdated or wrong.

Compile NYC landmarks, DOB, ACRIS, HPD, and BSA findings into one property due-diligence report. Use for a full report; use individual NYC skills for one dataset.

Python library for genomic interval ML. Train/apply region2vec embeddings turning BED regions into vectors, index interval datasets for ML, search embedding space with BEDSpace, and evaluate embedding quality. Use for chromatin accessibility clustering, regulatory element classification, and cross-sample region comparison.

Query CELLxGENE Census (61M+ cells). Search by cell type/tissue/disease/organism; get AnnData, stream out-of-core, train PyTorch models. For your own data use scanpy; for annotated data use anndata.

Query and download NCI Imaging Data Commons (IDC) cancer radiology and pathology datasets via the idc-index Python client. No authentication required: the parquet index ships inside the pip wheel, SQL runs locally via DuckDB, and DICOM downloads stream from public S3/GCS buckets through s5cmd. Use sql_query() for DuckDB cohort selection, get_collections/get_patients/get_dicom_studies/get_dicom_series for hierarchical browsing, download_from_selection() for downloads, and get_viewer_URL() for OHIF/Slim links. Use pydicom-medical-imaging for local DICOM reading; histolab for whole-slide pathology preprocessing.

Computational pathology toolkit for whole-slide images (WSIs): load slides, extract tiles, stain normalization, nuclear segmentation, feature extraction, and ML training. Supports H&E and multiplex. For end-to-end pipelines from raw WSIs to quantitative outputs.

Protein language models (ESM3, ESM C) for sequence generation, structure prediction, inverse folding, and embeddings. Design novel proteins, extract ML features, or fold sequences. Local GPU or EvolutionaryScale Forge API. Use AlphaFold for traditional folding; RDKit for small molecules.

Python library for healthcare ML on EHR data: process MIMIC-III/IV, eICU, OMOP-CDM; encode medical codes (ICD, ATC, NDC); build patient-level datasets; train Transformer, RETAIN, GRASP, MedBERT for mortality, drug recommendation, readmission, diagnosis prediction. Alternatives: FIDDLE (preprocessing), clinical-longformer (clinical NLP), ehr-ml (embeddings).

Structured hypothesis formulation: turn observations into testable hypotheses with predictions, propose mechanisms, design experiments. Follows the scientific method. Use scientific-brainstorming for open ideation; hypogenic for automated LLM hypothesis testing on datasets.

Deep learning for drug discovery. 60+ models (GCN, GAT, AttentiveFP, MPNN, ChemBERTa, GROVER), 50+ featurizers, MoleculeNet benchmarks, HPO, transfer learning. Unified load-featurize-split-train-evaluate API. For fingerprints use rdkit-cheminformatics; for featurization-only use molfeat.

Diffusion-based docking that predicts protein-ligand poses without a predefined site. Use for blind docking, when traditional docking fails, or exploring multiple binding modes. Pipeline: prep protein (PDB) and ligand (SMILES/SDF), run inference, analyze confidence-ranked poses.

Molecular featurization hub (100+ featurizers) for ML. SMILES to fingerprints (ECFP, MACCS, MAP4), descriptors (RDKit 2D, Mordred), pretrained embeddings (ChemBERTa, GIN, Graphormer), pharmacophores. Scikit-learn compatible with parallelization/caching. For QSAR, virtual screening, similarity, and molecular DL.

PyTorch-based ML platform for drug discovery: graph molecular representation learning, property prediction (ADMET, activity), retrosynthesis, drug-target interaction (DTI), and pretraining on large molecular datasets. Provides GNN layers (GraphConv, GAT, MPNN), pretrained models, and benchmark datasets.

Open-source FAIR biology data framework. Version artifacts (AnnData, DataFrame, Zarr), track lineage, validate via ontologies (Bionty), query datasets. Integrates with Nextflow, Snakemake, W&B, scVI. For scRNA-seq use scanpy; for ontology lookups use bionty.

Use for Kaggle competition overview fetches, writeups, discussion/kernel research, submissions, and dataset uploads. Not for unrelated ML code.

Use up-to-date library and framework docs via Context7 MCP instead of training data. Activates for setup questions, API references, code examples, or when the user names a framework (e.g. React, Next.js, Prisma).