skills

Backblaze B2 cloud storage via b2sdk Python SDK. Auth, upload, download, buckets, keys, large files, sync.

Docker on Mac with OrbStack. Containers, Dockerfiles, images, Compose, BuildKit, volumes, networking. Use when working with Dockerfiles, compose.yaml, docker/docker compose/buildx commands, or containerizing any service.

Electron desktop app development across macOS and Windows. Including security, performance, using code beyond Javascript & Typescript, native APIs, storage, debugging and troubleshooting issues, storage, and electron best practices. Use when working with electron code, troubleshooting, testing, preparing for release.

Build Electron apps for macOS with Next.js/React/TypeScript/Bun. Covers project setup, main/renderer/preload architecture, IPC, macOS window styling, vibrancy sidebar, traffic lights, native file system, dialogs, audio capture, desktopCapturer, storage (electron-store, SQLite), native modules (N-API, node-gyp, @electron/rebuild), memory optimization (V8 cage, leaks, window pooling), performance (startup, main-thread, virtualization), permissions and entitlements, notifications, dock/tray/menus, deep links, auto-update (electron-updater, Squirrel.Mac), electron-builder packaging, DMG, code signing, notarytool notarization, universal arm64+x64 builds, Mac App Store, context isolation and contextBridge security. Use when building, debugging, packaging, or distributing any Electron app targeting macOS.

Relevant when connecting your app/code with Google Calendar API v3 via TypeScript/JavaScript. Includes OAuth2, token refresh, CRUD on events, recurring events, conference/join links, RSVP, attendees, incremental sync, push notifications, and more.

When using the `datasets` library, `huggingface_hub`, or working with datasets on the HuggingFace Hub. Covers loading, streaming, resumable download, partial access, discovery, creation, upload, folder builders, WebDataset, data cards, caching, FAISS, and integrations.

Find & remove duplicate/near-duplicate images using perceptual hashing (PHash, AHash, DHash, WHash) and CNN embeddings. Use when deduplicating image datasets or comparing image similarity in python. Libraries: imagededup and imagehash

Build LLM evaluations with Inspect AI (inspect-ai Python package by UK AISI). Use this skill whenever the user mentions Inspect AI, inspect-ai, LLM evaluation frameworks, eval tasks, eval datasets, solvers, scorers, agent evals, model grading, agentic benchmarks, SWE-bench evals, or any task involving evaluating language models systematically. Also trigger for questions about sandboxing model code execution, tool-use evals, multi-agent evaluation, eval log analysis, or running benchmarks like MMLU, HumanEval, GSM8K, HellaSwag, ARC with Inspect. Even if the user just says "write an eval" or "benchmark this model", consider this skill.

Embedded vector database using LanceDB with TypeScript/JavaScript SDK. Use when building vector search, semantic search, RAG pipelines, hybrid search, full-text search, or multimodal data storage with LanceDB. Trigger on mentions of LanceDB, Lance format, vector similarity search with embedded DB, or vectordb that runs in-process. Also trigger when user needs an embedded alternative to Pinecone/Weaviate/Qdrant.

LM Studio Python SDK (lmstudio package). Chat completions, streaming, structured output, agentic tool use (.act()), embeddings, tokenization, model management, VLM image input. Use when writing Python code that talks to LM Studio, local LLMs via lmstudio SDK, or building agents with local models. Also covers lms CLI, llmster headless daemon, LM Link remote access, and server configuration.

LM Studio TypeScript SDK (@lmstudio/sdk), lmstudio-js, local LLM inference, chat completions, structured output, tool use, agentic .act(), plugins, embeddings, model loading, LM Link, lms CLI, llmster headless daemon

Capture system audio, microphone audio, and screenshots on macOS. ScreenCaptureKit, Core Audio Taps, AVAudioEngine, SCStream, SCStreamConfiguration, loopback audio, echo cancellation, noise suppression, voice processing, SCScreenshotManager, Electron desktopCapturer, AudioTee, aggregate device, CATapDescription, silence detection, screen lock persistence, CMSampleBuffer, AVAssetWriter, PCM audio streaming. Use when building meeting recorders, transcription apps, screen recording tools, or any app capturing system/mic audio on macOS.

Connect to MCP servers from TypeScript using Vercel AI SDK's @ai-sdk/mcp client. Use when building MCP clients, consuming MCP tools/resources/prompts, connecting to remote or local MCP servers, or integrating MCP with generateText/streamText. Covers HTTP, SSE, and stdio transports, schema discovery vs definition, typed outputs, resources, prompts, elicitation, and lifecycle management.

Build MCP servers in TypeScript using the official SDK. Exposing internal functions as tools, prompts, resources. Covers transport (stdio, Streamable HTTP), auth (API keys, OAuth 2.1), debugging with MCP Inspector, and remote exposure via Tailscale/Cloudflare Tunnel. Use when building, scaffolding, or deploying MCP servers in TypeScript/Node.

Next.js 16 App Router reference — routing, layouts, Server/Client Components, Cache Components (`use cache`), Server Actions, Route Handlers, `proxy.ts` (formerly middleware), styling, images, fonts, config, deployment. Use when building, reviewing, or debugging Next.js App Router code.

Build AI agents with PydanticAI — tools, structured output, multi-provider, streaming, multi-agent delegation, MCP, capabilities, evals, graphs. Use when working with pydantic_ai, pydantic-ai, Agent, RunContext, tool decorator, output_type, deps_type, capabilities, or any PydanticAI agent pattern in Python.

Pydantic v2 data validation, serialization, BaseModel, Field, validators, TypeAdapter, validate_call, ConfigDict, model_dump, model_validate, JSON Schema, custom types, SQLAlchemy integration, pydantic-settings, pydantic-extra-types. Use when writing Python models, schemas, API request/response types, config objects, CLI args, ORM layers, or any data validation. Also trigger when user has manual dict/JSON serialization that should use Pydantic, or raw dataclasses that need validation.

Custom and community GPU kernels for PyTorch transformer training and inference. Use when optimizing attention (FlashAttention, FlexAttention, SDPA, NATTEN, sliding window, paged, neighborhood), MLP fusion (Liger, SwiGLU, GeGLU, FusedMLP), normalization (RMSNorm, LayerNorm), RoPE, CrossEntropy, or optimizers (fused AdamW). Especially relevant for DiT, MMDiT, video DiT, S3DiT architectures. Also covers Colab A100 setup for flash-attn, torch.compile kernel fusion, and Triton kernel libraries like Unsloth and Liger Kernel.

Optimise PyTorch single-GPU training wall-clock time. torch.compile, CUDA graphs, Tensor Cores, AMP, profiling, memory pre-allocation, Triton kernels, compiled optimizers, gradient monitoring, model visualization, benchmarking. Use when training DiT, transformers, diffusion models on GPU and user mentions speed, throughput, compilation, graph breaks, reduce-overhead, profiling, nsight, kernel fusion, gradient flow, vanishing/exploding gradients, hooks, torchinfo, torchviz, record_function, benchmark, or training step timing. Also trigger for looped/recursive transformers with shared weights, weight tying, depth-wise LoRA, or routing in looped blocks.

RunPod GPU pods — Python SDK, REST API, CLI. Create/start/stop/terminate pods, spot instances, network volumes, templates, storage, SSH, file transfer, Backblaze B2 sync, Docker images, environment variables, secrets.

Build AI agents for Slack using Bolt for JavaScript/TypeScript. Use when building Slack bots with AI capabilities, assistant containers, agent streaming, suggested prompts, feedback buttons, App Home, modals, slash commands, @mentions, context management, or the Slack MCP Server. Also trigger for Slack CLI agent scaffolding, text streaming APIs, task cards, plan display, workspace search via assistant.search.context, or deploying Slack agents. Even if the user just says 'Slack bot with AI' or 'chat agent in Slack', use this skill.

Private mesh networking with Tailscale — server setup, auth keys, ACLs, Tailscale Serve/Funnel, Services (TailVIPs), firewall lockdown, cloud-init provisioning, and the tailscale Python API client. Use when connecting backend servers, workers, or services over Tailscale, setting up MagicDNS, HTTPS certs, or locking down VPS/cloud nodes.

Code-quality patterns and anti-patterns for TypeScript apps using React, TanStack Query, Zustand, Next.js / RSC. Use when writing new components or pages, reviewing existing code, setting up a new project, debugging re-renders, structuring data fetching, deciding between server and client components, designing UI states, choosing between WebSocket and HTTP, or anytime an `any` or `useEffect` is about to be added.

Rent GPU instances on Vast.ai marketplace using Python SDK. Search offers, create on-demand/interruptible instances, manage lifecycle, templates, volumes, data transfer, cloud sync. Use when mentioning vast.ai, VastAI, GPU rental, spot GPU, interruptible GPU instances, marketplace GPU.

Build WebSocket hubs with FastAPI, Pydantic, SQLite-backed channels, and in-memory fanout for a Mothership ↔ Workers/Frontends star topology over Tailscale. Use whenever the user mentions websockets, WS channels, real-time push, streaming status updates, pub/sub between mothership and remote workers, or connecting a React frontend via TanStack Query to a WS hub.

WorkOS AuthKit authentication — Next.js integration, sessions, RBAC roles/permissions, organizations, SSO, invitations, OAuth Connect, Actions webhooks, Stripe add-on. Use when adding auth with WorkOS, authkit-nextjs, or enterprise SSO/SCIM.

Animate React components with anime.js v4. Use when animating UI elements, creating motion/transitions, SVG animations, layout animations, or sequenced timelines in React. Covers the createScope pattern, animate(), timelines, SVG utilities, and the Layout API.

Add Clerk authentication to TypeScript/JS apps. Use when the user mentions Clerk, @clerk/nextjs, ClerkProvider, clerkMiddleware, or wants to add auth/protect routes/manage sessions in a Next.js project. Covers App Router and Pages Router setup, route protection, RBAC, session tokens, redirect URLs, CSP headers.

Typography system using Fira Sans. Apply when building web UIs, design systems, or any artifact requiring consistent type. Covers font loading, typescale, spacing rules, and HTML element mapping.

Google Gemini API integration using the @google/genai SDK for JavaScript/TypeScript. Covers text generation, image generation (Nano Banana), video generation (Veo), speech generation (TTS), image/video/audio/document understanding, thinking, structured outputs, function calling, embeddings, Live API (real-time streaming), batch processing, and safety settings. Use this skill whenever the user mentions Gemini, Google AI, @google/genai, Nano Banana, Veo, GoogleGenAI, or wants to integrate with Google's generative AI models in a JS/TS project — even if they just say 'call Gemini' or 'use Google AI'.

Integrate with the Gmail API for reading, sending, syncing, and managing email programmatically. Use when building email clients, automating email workflows, syncing mailboxes, processing attachments, managing labels/filters, or setting up push notifications via Pub/Sub. Covers OAuth 2.0 authentication (including localhost/dev and Cloudflare tunnels), full and partial sync, Pub/Sub push, search/query operators, batch requests, drafts, attachments, labels, filters, and performance best practices. Supports Python and Node.js SDKs.

Build applications using the xAI Grok API. Use when the user needs to call Grok models via API — text generation, reasoning, vision/image understanding, image generation, structured outputs, batch processing, streaming, function calling, or any xAI SDK integration. Covers both the Python xAI SDK (xai-sdk) and REST/TypeScript/JavaScript (OpenAI-compatible). Triggers on mentions of 'Grok API', 'xAI API', 'xai-sdk', 'grok-4', 'grok-3', or requests to integrate Grok into applications.

Work with HuggingFace datasets — loading, creating, uploading, streaming, partial downloading, and configuring data cards. Use this skill whenever the user mentions HuggingFace datasets, the `datasets` library, `load_dataset`, dataset repos on the Hub, data cards/README YAML configs, streaming large datasets, or partial/selective downloading of dataset files. Also trigger when the user wants to push data to the Hub, create ImageFolder/AudioFolder/VideoFolder datasets, use WebDataset format, or work with `huggingface_hub` for file-level dataset access. Trigger even if the user just says 'dataset' in the context of ML training pipelines — HuggingFace datasets is the de-facto standard.

React WYSIWYG markdown/MDX editor (@mdxeditor/editor). Use when building rich-text editors with markdown output, CMS editors, or any contentEditable markdown UI in React. Covers plugins, toolbar, theming, Tailwind Typography, code blocks, images, tables, JSX components, diff/source mode, and AI autocomplete.

Build WYSIWYG markdown editors with Milkdown, a plugin-driven framework built on ProseMirror and Remark. Use this skill whenever the user wants to create, customize, or integrate a rich-text markdown editor using Milkdown — whether via the batteries-included Crepe editor or the granular Kit/Core API. Also trigger when the user mentions Milkdown plugins, slash commands, toolbar creation, editor listeners, ProseMirror node views in a Milkdown context, @milkdown/react or @milkdown/crepe, or building WYSIWYG markdown editing into React/Vue/Svelte/vanilla apps. Covers getting started, plugin usage, React integration, keyboard shortcuts, styling, and the slash-plugin example. Even if the user just says 'markdown editor' or 'WYSIWYG editor' alongside mentions of ProseMirror or Remark, consider whether Milkdown is the right tool.

Build third-party API integrations using Nango — managed OAuth, authenticated proxy requests, and continuous data syncs from 600+ APIs. Use when connecting to Slack, Gmail, Google Docs, Google Calendar, Google Drive, Linear, PostHog, Metabase, or any external API where you need user auth flows, token management, data syncing, or proxy requests. Covers Auth (OAuth 2.0, API keys), Proxy (authenticated pass-through), Functions (syncs + actions), and the AI integration builder.

SOLID principles for Swift 6 and SwiftUI (iOS 26+). Files < 100 lines, protocols separated, @Observable, actors, Preview-driven development. Features Modular MANDATORY.

Typography system using Space Grotesk. Apply when building web UIs, design systems, or any artifact requiring consistent type. Covers font loading, typescale, spacing rules, and HTML element mapping.

Comprehensively reviews SwiftUI code for best practices on modern APIs, maintainability, and performance. Use when reading, writing, or reviewing SwiftUI projects.

Build cross-platform desktop and mobile applications with Tauri v2, Rust backend, and web frontends (React/Next.js, Vue, Svelte, etc.). Use this skill whenever the user wants to create, scaffold, develop, debug, or distribute a Tauri application. Also trigger when the user mentions Tauri plugins, Tauri commands, Tauri capabilities/permissions, calling Rust from JavaScript, IPC between frontend and backend, native desktop/mobile app packaging, or building apps with pnpm/bun + Rust. Covers project setup, Next.js SSG integration, plugin installation, security capabilities, code signing, and distribution to macOS App Store, Google Play, Windows Installer, AppImage, and more. Even if the user just says 'build a desktop app' or 'package my web app as native', consider whether Tauri is the right tool and suggest this skill.