Machine Learning Skills

LLM and AI application security testing skill for prompt injection, jailbreaking, and AI system vulnerabilities. This skill should be used when testing AI/ML applications for security issues, performing prompt injection attacks, testing LLM guardrails, analyzing AI system architectures for vulnerabilities, or assessing RAG pipeline security. Triggers on requests to test LLM security, perform prompt injection, jailbreak AI systems, test AI guardrails, or audit AI application security.

This skill should be used when the user asks to "fine-tune on books", "create SFT dataset", "train style model", "extract ePub text", or mentions style transfer, LoRA training, book segmentation, or author voice replication.

Execute Hugging Face Hub operations using the `hf` CLI. Use when the user needs to download models/datasets/spaces, upload files to Hub repositories, create repos, manage local cache, or run comput...

| issue, the version-bump and changelog prep PR, or the post-release development-version bump PR. Reads release metadata from `<project-config>/release-trains.md` and `<project-config>/release-management-config.md`. Every output is a draft confirmed by the Release Manager before filing; the agent never marks a PR ready, never merges, and never closes any artefact.

Compose Mapbox MCP tools to produce grounded, cited location-aware responses from live data instead of training data

> Marketing benchmark triage skill. Use whenever metrics are shared with questions like "is this good?", "is this normal?", or "should I worry?" — ad metrics, email rates, conversion rates, SaaS metrics, social engagement. Compares the user's numbers against published, sourced benchmarks using panel-matching rules (same platform, model, market, and denominator), abnormal and which are benchmark anxiety.

> Use the moment a request is underspecified and you are about to act on an ASSUMED goal. Enumerate the candidate goals the request could mean, project two or more plausible goals lead to materially different actions or artifacts, ask exactly ONE targeted clarifying question; otherwise proceed on the most-likely goal and state the assumption in one line. Clarification is an evidence-producing action, not a delay — but one question, never an interrogation. Distinct from gsd-spec-phase (GSD-phase-bound, emits a heavy SPEC.md) and intent-router (fetch strategy, not goal disambiguation). Backed by agent goal-state inference (arxiv 2606.16813v1). Triggers on acting under an assumed goal when the request admits more than one materially different reading.

> Use at write time to vector/embedding memory — Grove content-addressed insertion, chroma/pgvector upserts, memory-consolidation promoting session traces to MEMORY.md, or embedding externally-ingested content. Scores each candidate record against a fixed panel of sentinel queries and quarantines any record that would become the nearest neighbor of too many unrelated queries — a hub — whether from adversarial poisoning or accidental over-generality. This is the memory-record-side sibling of skill-injection-guardian (file-side) and the write-side complement of memory-use-warrant (read-side). Quarantine, never silently drop; a human reviews. Backed by the admission-time hubness gate (arxiv 2606.19692v1). Triggers on inserting into vector memory, consolidating memory, and embedding stranger content.

Complex systems and emergent behavior — chaos theory, fractals, neural networks, cellular automata, and computability. Analyzes dynamical systems for chaos via Lyapunov exponents, classifies bifurcations, measures fractal dimension, and models emergent phenomena. Use when analyzing chaotic systems, computing Lyapunov exponents, classifying bifurcations, measuring fractal dimension, modeling neural networks, or studying emergent behavior from simple rules.

Sketched Isotropic Gaussian Regularization primitive. Scalar loss matching the embedding distribution to a standard-normal target via Cramér-Wold slicing and the Epps-Pulley empirical characteristic function test. Port of rbalestr-lab/lejepa (MIT). Default-off in v1.49.571.

Framework for building competitive landscape decks — market positioning, competitor deep-dives, comparative analysis, strategic synthesis. Use when the user asks for a competitive landscape, competitor analysis, peer comparison, market positioning assessment, strategic review, or investment memo deck. Also triggers on "who are the competitors to X", "benchmark X against peers", "build a market map", or any request to systematically evaluate competitive dynamics across an industry.

Use whenever the user wants to find, shortlist, vet, or enrich US AI/ML/data consulting firms (consultancies) — AI/ML development, MLOps, generative AI / LLM apps (RAG, chatbots, agents), computer vision, NLP, recommendation systems, data engineering, BI/analytics. Triggers on "find an AI/ML consulting firm to build our recommendation engine", "shortlist three RAG/LLM consultancies for an enterprise chatbot", "compare three AI/ML consulting firms with strong ratings", or "pull contact info for these 8 AI consultancy domains", even when described indirectly (we want to use AI for X, deploy ML to production). Drives the ServiceGraph API (api.servicegraph.co) — a 100k+ US firm catalog filterable by industry, services, location, size, ratings. Defer to find-software-developer for general app/backend work where AI is just a feature. Skip in-house ML/data hires, LLM/AI-tool comparisons (ChatGPT vs Claude), "how do I fine-tune X" DIY questions, AI courses for individuals, non-US firms, individual freelancers.

Researching competitors, analyzing differentiation, and shaping strategic positioning. Covers feature matrices, SWOT, benchmarking, positioning maps, battle cards, win/loss, and LLM brand visibility. Research only — no code. Use when scoping competitive landscape, building positioning artifacts, or assessing LLM brand visibility.

Designing and evaluating AI/ML systems across prompt engineering, RAG design, LLM application patterns, AI safety, evaluation frameworks, MLOps, and cost optimization. Use when designing AI/ML pipelines, RAG architectures, prompt strategies, evaluation harnesses, or LLM cost models.

Designing search engines and vector DBs for full-text, vector, and hybrid retrieval. Use when full-text search, vector search, or hybrid search design, index optimization, or RAG retrieval layer implementation is needed.

Patterns for Retrieval-Augmented Generation (RAG) and agent memory systems. Retrieves only relevant context, prevents context bloat, and maintains coherent state across sessions.

Measure first, optimize second. Data-driven performance improvements with before/after benchmarks and production validation.

Autonomously research, implement, train and ship ML code using the Hugging Face ecosystem. Port of huggingface/ml-intern as a Claude Code skill. Triggers when the user asks to implement, train, fine-tune, or reproduce an ML model / paper / dataset workflow (e.g. "implement DeepSeek-V3 at 100M", "fine-tune Qwen on dataset X", "reproduce paper Y"). Clarifies ambiguous tasks before starting, runs under an explicit experiment budget, explores multiple viable solution paths in parallel via implementation subagents, and diagnoses + retries failed runs. HF-native: pulls datasets/models/papers from the Hub, pushes trained checkpoints + run logs back to the Hub. Emits Telegram + Slack milestone alerts via scripts/notify.sh.

Local LLM operations with Ollama on Apple Silicon, including setup, model pulls, chat launchers, benchmarks, and diagnostics.

LangChain LLM application framework with chains, agents, RAG, and memory for building AI-powered applications

AI session compression techniques for managing multi-turn conversations efficiently through summarization, embedding-based retrieval, and intelligent context management.

> Detect performance anti-patterns and apply optimization techniques in Go. Covers allocations, string handling, slice/map preallocation, sync.Pool, benchmarking, and profiling with pprof. Use when checking performance, finding slow code, reducing allocations, profiling, or reviewing hot paths. "benchmark this", "profile", "optimize Go code". Do NOT use for concurrency correctness (use go-concurrency-review) or general code style (use go-coding-standards).

| Use when a user runs, debugs, verifies, or ships a DL experiment on a GPU they OWN or RENT (AutoDL, RunPod, vast.ai, Lambda, Paperspace, 恒源云/矩池云/Featurize/揽睿星舟, bare SSH, Slurm, K8s; spot 抢占/preemption, 断点续训/resumable, 防 SSH 断线/tmux 守护, 多实例 ablation, 关机/销毁/stop-vs-terminate billing, checkpoint 磁盘满, CUDA OOM/显存不足, loss NaN/spike/不收敛, overfit 单 batch, FSDP/DeepSpeed/torchrun, 多卡 hang, 训练太慢/GPU util 低, dataloader/数据增广 bug; 消融结果异常/ablation looks wrong, 复现/reproducibility, 数据泄漏/leakage/test-set tuning, mAP=0/全零指标, 输出恒定/model-ignores-input, train-good/val-collapse, 对比不公平/unfair baseline, 单 seed/no error bars, loss 太好/too-good-to-be-true, 跨文档对账/cross-doc drift; 交付产物/deliverable, 唯一真源/single source of truth, best ckpt 拉回, 结果可视化/论文图脚本, manifest/provenance, 一键复现/repro, EVIDENCE.json. NOT for multi-cloud price-shopping + auto spot-recovery (SkyPilot), BYOC dev environments (dstack), or zero-ops serverless inference (Modal).

Use this skill when reviewing, designing, or improving a workshop, lesson, training session, facilitation plan, talk, or educational explanation. Use it when the user provides a topic, concept, or explanation and wants better framing questions, curiosity hooks, opening prompts, discussion questions, learner reflection prompts, or ways to make participants think before being taught.

Verifies claims in backlog items, skill documentation, or plugin content against primary sources using web lookups. Spawns parallel verification agents that must use WebFetch/WebSearch/gh — training data recall is explicitly rejected as evidence. Produces VERIFIED/REFUTED/INCONCLUSIVE verdicts with citations. Use when items are marked UNVERIFIED or when verifying tool API claims, CLI flags, or documented software behavior.

When setting up local LLM inference without cloud APIs. When running GGUF models locally. When needing OpenAI-compatible API from a local model. When building offline/air-gapped AI tools. When troubleshooting local LLM server connections.

Audit skill lifecycle by tracing call chains, detecting circular dependencies, finding instruction contradictions, identifying duplicated datasets, analyzing bidirectional coherence, discovering scriptable sequences, and learning patterns. Use when checking skill coherence, validating skill workflow, finding semantic gaps in plugin structure, or auditing plugin before marketplace submission. Generates audit reports to .plugin-creator/audits/ with findings by dimension.

> Plan verification and validation campaigns for simulation codes using manufactured solutions, canonical benchmark problems, grid/time refinement, uncertainty propagation, and pass/fail acceptance criteria. Use when an agent needs to prove a solver, model, or result is trustworthy rather than only plausible.

| R performance profiling, benchmarking, and optimization strategies. Use this skill when code is running slowly, comparing alternative implementations, deciding between dplyr/data.table/base R, or implementing parallel processing. Covers profvis and bench usage, performance workflow, parallel processing with in_parallel(), data backend selection, modern purrr patterns (list_rbind, walk), and common performance anti-patterns to avoid.

Implements the Syncfusion .NET MAUI Range Selector (SfRangeSelector) component. Use when working with range selection with embedded charts, numeric range input, RangeStart/RangeEnd, track customization, or thumb styling in .NET MAUI. Covers tooltip configuration, label formatting, ticks/dividers, selection modes, events, and embedding SfCartesianChart inside a range control.

Use when an award run needs external evidence — literature, datasets, benchmarks, domain constants, prior-art checks, or citation verification — with combined built-in WebSearch and search MCP discovery, source reconciliation, document reading, citation confirmation, token budgets, and reproducibility rules.

| Before declaring any task complete, actually verify the outcome. Run the code. Test the fix. Check the output. Claude's training optimizes for plausible-looking output, not verified-correct output. This skill forces the verification step that doesn't come naturally. No victory laps without proof.

Use when migrating from SwiftData to SQLiteData — decision guide, pattern equivalents, code examples, CloudKit sharing (SwiftData can't), performance benchmarks, gradual migration strategy

Use when wrapping UIKit views/controllers in SwiftUI, embedding SwiftUI in UIKit, or debugging UIKit-SwiftUI interop issues. Covers UIViewRepresentable, UIViewControllerRepresentable, UIHostingController, UIHostingConfiguration, coordinators, lifecycle, state binding, memory management.

CoreML diagnostics - model load failures, slow inference, memory issues, compression accuracy loss, compute unit problems, conversion errors.

Use when deploying custom ML models on-device, converting PyTorch models, compressing models, implementing LLM inference, or optimizing CoreML performance. Covers model conversion, compression, stateful models, KV-cache, multi-function models, MLTensor.

Semantic Similarity Index for disease research literature using PubMedBERT embeddings

Local scVI-based single-cell latent embedding and batch-aware integration from raw-count .h5ad or 10x Matrix Market input, with stable integrated AnnData export for downstream latent analysis.

Automate Benchmark Email tasks via Rube MCP (Composio). Always search tools first for current schemas.

Automate Google BigQuery tasks via Rube MCP (Composio): run SQL queries, explore datasets and metadata, execute MBQL queries via Metabase integration. Always search tools first for current schemas.

Automate Mistral AI operations -- manage files and libraries, upload documents for fine-tuning, batch processing, and OCR, track fine-tuning jobs, and build RAG pipelines via the Composio MCP integration.

Automate Mistral AI tasks via Rube MCP (Composio): completions, embeddings, fine-tuning, and model management. Always search tools first for current schemas.

Automate OpenAI API operations -- generate responses with multimodal and structured output support, create embeddings, generate images, and list models via the Composio MCP integration.

Golang benchmarking, profiling, and performance measurement. Use when writing, running, or comparing Go benchmarks, profiling hot paths with pprof, interpreting CPU/memory/trace profiles, analyzing results with benchstat, setting up CI benchmark regression detection, or investigating production performance with Prometheus runtime metrics. Also use when the developer needs deep analysis on a specific performance indicator - this skill provides the measurement methodology, while golang-performance provides the optimization patterns.

INVOKE THIS SKILL when creating evaluation datasets, uploading datasets to LangSmith, or managing existing datasets. Covers dataset types (final_response, single_step, trajectory, RAG), CLI management commands, SDK-based creation, and example management. Uses the langsmith CLI tool.

Standards for Baseline Profiles, Startup Time, and UI Rendering. Use when optimizing app startup, jank, or UI rendering performance in Android. (triggers: **/*Benchmark.kt, **/*Initializer.kt, BaselineProfile, JankStats, recomposition)