Machine Learning Skills

> Structure prediction using OpenFold3, an open-weights PyTorch reproduction of AlphaFold3 from the AlQuraishi Lab. Use this skill when predicting protein/nucleic-acid/ligand complex structures with an Apache-2.0-licensed AF3 reimplementation.

> Embed and annotate single-cell expression data with scGPT, a foundation model (1) Producing cell embeddings from an AnnData for clustering/integration, (2) Zero-shot or fine-tuned cell-type annotation, (3) Gene-level representation for perturbation/GRN tasks. For probabilistic single-cell models (scVI etc.), use the scvi-tools library.

Invoke an already configured model endpoint from a supported Wisp execution context and capture the bounded inference as a Run. Use only when the endpoint URL and authentication are already available inside that context; this skill does not register or manage services.

Validates dataset formatting and quality for SageMaker model fine-tuning (SFT, DPO, or RLVR). Use when the user says "is my dataset okay", "evaluate my data", "check my training data", "I have my own data", or before starting any fine-tuning job. Detects file format, checks schema compliance against the selected model and technique, and reports whether the data is ready for training or evaluation.

Generates code that transforms datasets between ML schemas for model training or evaluation. Use when the user says "transform", "convert", "reformat", "change the format", or when a dataset's schema needs to change to match the target format — always use this skill for format changes rather than writing inline transformation code. Supports OpenAI chat, SageMaker SFT/DPO/RLVR/RLAIF, HuggingFace preference, Bedrock Nova, VERL, and custom JSONL formats from local files or S3.

Selects a fine-tuning technique (SFT, DPO, RLVR, or RLAIF) for the user's use case and validates it against the selected model's available recipes. Use when the user has decided to finetune and needs to choose a technique, or when the technique needs to be validated against a model. Requires a base model to already be selected (via model-selection skill).

Generates code that fine-tunes a base model using SageMaker serverless training jobs. Use when the user says "start training", "fine-tune my model", "I'm ready to train", or when the plan reaches the finetuning step. Supports SFT, DPO, RLVR, and RLAIF trainers, including RLVR Lambda reward function and RLAIF custom prompt creation.

Diagnose NCCL failures and adjacent training-pod failures on HyperPod GPU clusters (EKS or Slurm) — training hangs, AllReduce / collective-op timeouts, EFA or libfabric errors, rendezvous failures, EFA TCP fallback, /dev/shm or memlock issues, NCCL version mismatch across pods, container OOM / exit-137 / OOMKilled, GPU OOM (CUDA out of memory), CrashLoopBackOff / Pending pods, MASTER_ADDR DNS, NetworkPolicy blocking. Not for single-node hardware faults (→ hyperpod-node-debugger § G) or cluster-creation EFA / SSM failures (→ hyperpod-cluster-debugger § A / § F).

Check and compare software component versions on SageMaker HyperPod cluster nodes - NVIDIA drivers, CUDA toolkit, cuDNN, NCCL, EFA, AWS OFI NCCL, GDRCopy, MPI, Neuron SDK (Trainium/Inferentia), Python, and PyTorch. Use when checking component versions, verifying CUDA/driver compatibility, detecting version mismatches across nodes, planning upgrades, documenting cluster configuration, or troubleshooting version-related issues on HyperPod. Triggers on requests about versions, compatibility, component checks, or upgrade planning for HyperPod clusters.

Generates python code that evaluates SageMaker models. Supports two evaluation types: LLM-as-Judge and Custom Scorer. Use when the user says "evaluate my model", "run a benchmark", "test model performance", "how did my model perform", "compare models", or other similar requests.

Selects a base model for the user's use case by querying SageMaker Hub. Use when the user asks which model to use, wants to select or change their base model, mentions a model name or family (e.g., "Llama", "Mistral", "Nova"), or wants to evaluate a base model — always activate even for known model names because the exact Hub model ID must be resolved. Queries available models, presents benchmarks and licenses, and confirms selection.

Discovers user intent and generates a structured, step-by-step plan for model customization workflows. This skill must always be activated alongside any other skill when the user's request relates to model customization — including fine-tuning, training, building, customizing, reviewing data, or getting advice on approach, regardless of domain. Do not skip this skill even if the immediate ask is narrow (e.g., reviewing data format or a single workflow step), because planning discovers the full scope of work needed. Also activate when the user wants to resume, continue, or modify an existing plan.

When the user wants to add or optimize structured data (Schema.org, JSON-LD). Also use when the user mentions "schema," "structured data," "JSON-LD," "rich results," "rich snippets," "Google rich snippets," "featured snippet schema," "add schema to page," "missing structured data," "schema validation error," "Schema Markup Validator," "Google Rich Results Test," "FAQ schema," "Article schema," "Organization schema," "JobPosting," "HowTo," "Event," "SoftwareApplication," "BreadcrumbList," "WebSite," "Recipe," "Product," or "Dataset." For SERP feature types and zero-click patterns, use serp-features. For AI search visibility strategy (not markup), use generative-engine-optimization. For HowTo step sections (placement, copy, vs FAQ), use howto-section-generator.

Run or resume a measured improvement loop in a local project. Use for benchmark-driven optimization, qualitative quality-gap research, packet logging, dashboard readouts, recovery, and review-branch finalization backed by autoresearch session files.

| Use when adding, debugging, or modifying LLMObs plugins for an LLM library completions / streaming / embeddings / agent runs / orchestration / tool calls / retrieval", "LLMObsPlugin", "getLLMObsSpanRegisterOptions", "setLLMObsTags", "SPAN_KINDS", "span kind", any provider tag ("openai" / "anthropic" / "genai" / "google" / "langchain" / "langgraph" / "ai" llmobs), "VCR cassettes".

>- Core information for AlphaEarth Foundations Satellite Embeddings in Google Earth Engine (GEE).

>- Use AlphaEarth Foundations Satellite Embeddings in Google Earth Engine (GEE) for change detection.

Integrate TileGym kernels into Hugging Face `transformers` models by replacing the library's submodule(s) and certain class(es)' implementations, and patching certain class(es)' init/forward/load weight methods prior to instantiating models. Used when the user requires integrating TileGym kernels into `transformers` models.

Add a new cuTile GPU kernel operator to TileGym. Covers dispatch registration in ops.py, cuTile backend implementation, __init__.py exports, test creation, and benchmark in tests/benchmark. Use when adding, creating, or implementing a new cuTile operator/kernel in TileGym, or when asking how to register a new cuTile op.

Iteratively optimize cuTile kernel performance through systematic profiling, bottleneck analysis, IR comparison, and targeted tuning. Covers tile sizes, occupancy, autotune configs, TMA, latency hints, persistent scheduling, num_ctas, flush_to_zero, and IR-level debugging. Use when asked to "optimize cutile kernel", "improve kernel perf", "tune cutile performance", "make kernel faster", or iteratively benchmark and refine a cuTile GPU kernel in the TileGym project.

> Use when the user wants to build an evaluation system for an LLM/agent application but doesn't know where to start — they have traces, prompts, RAG pipelines, or nothing at all. Also use when the user mentions evaluation, eval, benchmarking, testing LLM quality, measuring agent performance, assessing RAG accuracy, or questions then recommends which sub-skill (local workflow) to use next.

> Use when the user needs to design evaluation datasets, create test cases, stratify samples, generate adversarial examples, extract eval dimensions from traces/specs, or build a labeled evaluation set. Also use when the user mentions test data design, eval coverage, difficulty stratification, synthetic data generation for eval, or "how to create good evaluation data." Outputs datasets in OpenJudge-compatible format.

> Use when the user has evaluation principles or a dataset but needs help choosing the right graders, designing evaluation metrics, creating LLM-as-judge prompts, combining multiple metrics into a composite score, or building an automated evaluation pipeline. Also use when the user mentions grader selection, metric design, judge prompt engineering, rubric design, evaluation pipeline code, or "how to evaluate [X] automatically." Outputs executable OpenJudge pipeline code.

> Use when the user has a RAG (Retrieval-Augmented Generation) system and wants to evaluate its quality — separating retrieval issues from generation issues. Also use when the user mentions RAG evaluation, faithfulness checking, hallucination detection in RAG, retrieval quality, chunking optimization, or "is my RAG pipeline working." Outputs a diagnostic matrix that pinpoints whether problems are in retrieval or generation.

> Build RL reward signals using the OpenJudge framework. Covers choosing between pointwise and pairwise reward strategies based on RL algorithm, task type, and cost; aggregating multi-dimensional pointwise scores into a scalar reward; pairwise tournament reward for GRPO on subjective tasks (net win rate across group rollouts); generating preference pairs for DPO/RLAIF; and normalizing scores for training stability. Use when building reward models, scoring rollouts for GRPO/REINFORCE, generating preference data for DPO, or doing Best-of-N selection.

Use when the user wants to optimize configurable system parameters against a measurable scalar objective, especially for model training, inference, quantitative strategies, reinforcement learning, scientific workflows, or other expensive black-box evaluations where reading the project can improve trial selection.

Find and verify award-winning designs in the same or adjacent functional category through eight explicit relevance dimensions: problem and user, core function, sensing technology, intervention mechanism, physical form, use context and workflow, system architecture, and visual language. Use when a user asks for same-category winners, comparable precedents, design benchmarks, appearance-related award winners, or examples from iF Design, Red Dot, IDEA, or iF Design Student Award. Do not use this skill to score, judge, optimize, or match the user's design to an award.

用于对抗样本、模型提取、提示注入、成员推断、训练投毒、LoRA 滥用、LLM 越狱等 AI/ML 相关 CTF 题;触发名:ctf-ai-ml

Implement and extend PostHog Data warehouse import sources. Use when adding a new source under products/warehouse_sources/backend/temporal/data_imports/sources, adding datasets/endpoints to an existing source, or adding incremental sync, resumable imports, webhook ingestion, pagination, credentials validation, and source tests.

Expert knowledge for Azure Open Datasets development including limits & quotas. Use when handling non-Spark dataset downloads, throttling behavior, quota limits, retry logic, or rate-limit workarounds, and other Azure Open Datasets related development tasks. Not for Azure Data Explorer (use azure-data-explorer), Azure Synapse Analytics (use azure-synapse-analytics), Azure Databricks (use azure-databricks), Azure Machine Learning (use azure-machine-learning).

Expert knowledge for Microsoft Foundry Local (aka Azure AI Foundry Local) development including best practices, configuration, and integrations & coding patterns. Use when compiling HF models with Olive, using Foundry Local CLI, chat/embeddings APIs, transcription, or tool calling, and other Microsoft Foundry Local related development tasks. Not for Microsoft Foundry (use microsoft-foundry), Microsoft Foundry Classic (use microsoft-foundry-classic), Microsoft Foundry Tools (use microsoft-foundry-tools), Azure Local (use azure-local).

雪峰式AI-Native产品开发方法论。适用于：(1) 用户行为开放、不可穷举的AI-native产品（AI日历、AI助手、AI推荐、对话式产品等），(2) 强模型依赖型场景，AI驱动核心决策而非仅辅助，(3) 多专精Agent架构设计与分工，(4) 上线后快速校准、行为审计与漂移检测，(5) 模型选择和智能路由策略，(6) 概率性输出的质量评估。触发场景包括"AI-native产品怎么做"、"用户行为不可预测怎么办"、"多agent怎么分工"、"模型漂移怎么处理"、"校准到95%太难了"、"唯快不破"、"怎么选模型"、"agent并行分工"、"AI产品上线后怎么迭代"。注意：如果产品是场景明确、边界可定义的+AI类型，请改用 keqian-method skill。即使用户没有明确说"AI-native"，但在讨论AI驱动决策、用户行为不可预测、概率性输出等话题时也应触发。

> Run performance benchmarks to measure the impact of code changes. Discovers relevant benchmark scenarios based on changed files, executes them comparing a baseline version against local changes, and summarizes performance results. Use this when touching performance-sensitive code paths or when asked about performance impact.

Assess compliance with the EU General-Purpose AI (GPAI) Code of Practice under the AI Act (Regulation (EU) 2024/1689, KI-Verordnung) Articles 51-56. Covers GPAI model identification (Art. 3(63)), systemic risk model designation (Art. 3(65), Art. 51, 10^25 FLOP training compute threshold), upstream provider obligations under Art. 53 (transparency, technical documentation, copyright policy and EU rights reservation, training data summary template), additional systemic risk obligations under Art. 55 (model evaluation benchmarks, adversarial red-teaming, AI Office incident notification under Art. 55(1)(c), cybersecurity), AI Office notification (Art. 52), downstream provider duties (Art. 25 quasi-provider), and Art. 53(2) open-source LLM exemption. Maps the 12 GPAI Code of Practice commitments across transparency, copyright, and safety-and-security chapters. DACH: BaFin, BSI, BNetzA, Betriebsrat. Use when asked about GPAI provider obligations, LLM or foundation model compliance, the Art. 51 systemic risk threshold, Art. 53 transparency requirements, Art. 55 systemic-risk obligations, AI Office notification, downstream Art. 25 duties, Art. 53(2) open-source exemption, or generative AI provider compliance under the EU AI Act.

Routes any legal task to the right LLM, like OpenRouter but for legal work and grounded in benchmarks instead of brand loyalty. Built from mid-2026 legal evals (legalbenchmarks.ai, Vals AI × Stanford LegalBench across 124 models, Harvey's Legal Agent Benchmark, the Atticus Project's CUAD/MAUD/ACORD) plus translation evidence (WMT25, SwiLTra-Bench, ArabLegalEval). Covers five verticals: contract drafting, info extraction, legal research, contract review, and legal translation (including Arabic/MENA). Each asks up to four questions (cost, speed, accuracy/stakes, privacy/jurisdiction/language), then returns a primary model, a fallback, what to avoid, and what a human must verify. Core principle: capability is not controllability, so every route ends with a verification step. Not legal advice; a lawyer owns the output.

> Pick the right LLM for CONTRACT DRAFTING — generating, redlining, or rewriting contract language from instructions. Vendor-neutral routing grounded in mid-2026 legal benchmarks (legalbenchmarks.ai Contract Drafting). Asks up to 4 quick questions (cost, speed, accuracy/ stakes, privacy/jurisdiction/language), then recommends a primary model + fallback + what to avoid + what a human must verify. Use when someone asks "which model should I use to draft this clause/agreement", "best AI for drafting contracts", "route this drafting task", or is about to generate/redline contract text and hasn't fixed a model.

> Pick the right LLM for CONTRACT REVIEW — reviewing an existing agreement for risk, deviations from standard, missing protections, and internal contradictions, and proposing redlines. Vendor-neutral routing grounded in mid-2026 legal benchmarks. Contract review is a COMPOSITE skill (extraction + reasoning + redlining + conflict-detection), so this routes by the dominant sub-task. Asks up to 4 quick questions (cost, speed, accuracy/stakes, privacy/jurisdiction/ language), then recommends a primary model + fallback + what to avoid + what a human must verify. Use when someone asks "which model to review this contract / NDA / MSA", "best AI to redline", "route this contract review", or is about to review an agreement without a fixed model.

> Pick the right LLM for LEGAL INFO EXTRACTION — pulling facts, clauses, dates, parties, obligations, and structured fields out of contracts and legal documents. Vendor-neutral routing grounded in mid-2026 benchmarks (legalbenchmarks.ai Info Extraction; CUAD/MAUD/ACORD). Asks up to 4 quick questions (cost, speed, accuracy/stakes, privacy/jurisdiction/language), then recommends a primary model + fallback + what to avoid + what a human must verify. Use when someone asks "which model should I use to extract clauses/data from these documents", "best AI for contract data extraction", "route this extraction task", or is about to pull structured fields from legal docs without a fixed model.

Framework for demonstrating AI capabilities in legal contexts. Provides detailed personas across tenant law, business contracts, startup disputes, employment claims, and consumer protection with progressive complexity scenarios. Use when: (1) Demonstrating AI-powered legal triage or intake systems, (2) Showcasing responsible AI-assisted client interactions, (3) Training staff on appropriate AI use in legal contexts, (4) Creating realistic scenarios for legal tech presentations, (5) Developing educational materials about AI in legal services, or (6) Testing AI-powered legal information systems in controlled environments.

> Builds a high-fidelity interactive legal assessment as a single self-contained HTML artifact. Output includes a live countdown timer, contract review tasks with hover-annotated problem clauses, candidate answer textareas, model answers hidden behind reveal blocks, scenario-based legal memo tasks, strategy and function-building questions, and a pre-submission checklist that encodes the marking criteria. Use when the user needs to (1) assess a legal candidate with a realistic timed exercise, (2) train or onboard junior lawyers using problem sets rather than doctrine, (3) help a candidate prepare for a real take-home assessment they are facing, (4) build educational materials for law students, in-house teams, or compliance training, or (5) produce scenario-based training modules on specific legal topics. Triggers on "legal test", "take-home", "mock exam", "contract redline exercise", "candidate assessment", "legal training exercise", "practice test", or similar phrasing even when informal.

>- Act as experienced opposing counsel to attack, undermine, and expose weaknesses in a legal argument, submission, witness statement, or structured reasoning. 1. A core theory of attack identifying the single most effective way to defeat the argument; 2. A reconstructed version of the opposing argument stripped of rhetoric to expose its fragility; 3. Primary lines of attack grouped by category (legal misstatement, evidential gaps, causation failures, internal inconsistency, over-reliance on assertion, procedural weakness); 4. An "if I were the judge" section showing how a sceptical tribunal would dismantle the argument; 5. Surgical strikes - 3 to 5 high-impact points ready for oral submissions; and 6. An analysis of what the argument is trying to hide. Written in formal, adversarial British English for a legally trained audience.

Generates standalone interactive HTML "deal cards" that translate complex regulations into negotiation-ready reference tools, systematically distinguishing mandatory obligations from negotiable implementation choices. Use when the user needs an interactive regulatory guide for (1) contract negotiation support, (2) client education or internal training, (3) regulatory briefings for commercial stakeholders, or (4) structured comparison between required and flexible compliance paths. Primary focus on EU digital regulation (Data Act, AI Act, CRA, DORA, NIS2, GDPR) but the structural pattern transfers to any regulation where separating hard obligations from implementation choice is the point. Supports bilingual output where the jurisdiction calls for it.

Draft matter status reports from emails, call notes, and updates. Internal and client-facing formats, RAG logic, variance commentary, escalation flags. Use when asked to draft a status report, write a project update, summarise matter progress, prepare a client report, create a weekly or monthly update, convert emails into a status summary, or produce any kind of matter reporting. Also triggers when the user pastes email threads and asks what the status is, or needs to turn internal updates into client-facing reports.

Use when building AI agents in Dart, implementing Genkit flows or tools, integrating LLMs into Dart or Flutter applications, or using Genkit Dart plugins.

>- Inspects and configures the security headers a Power Pages site sends to browsers — Content Security Policy, frame and clickjacking protection, cross-origin sharing, cookie behavior, and related site settings. Identifies gaps and walks the user through fixes. Use when the user wants to review headers, fix CSP errors, allow embedding in another site, control cross-origin access, harden cookie settings, or asks "are my browser settings safe?", "fix my CSP", "set up CORS" — even if they only mention a specific header name without saying "security headers".

Swift 6.2 concurrency updates including default MainActor inference, @concurrent for background work, isolated conformances, and approachable concurrency migration. Use when adopting Swift 6.2 concurrency features or fixing data-race errors.

SwiftUI framework skills — data flow (identity, Observation, state ownership), layout & containers (Layout protocol, lazy-stack performance), AlarmKit, 3D charts, rich text editing, customizable toolbars, and WebKit embedding. Use for SwiftUI state/rendering bugs, custom layouts, scroll performance, or these feature areas.

WebKit integration in SwiftUI using WebView and WebPage for embedding web content, navigation, JavaScript interop, and customization. Use when embedding web content in SwiftUI apps.