Skills
97,905 packaged capabilities and playbooks for agents. Install any of them with one command.
Development
Use when a domain term needs defining, a glossary term is being used inconsistently or fuzzily, a hard-to-reverse decision needs an ADR, or existing ADRs need auditing, pruning, or archiving — produces glossary updates and a classified docs/adr tree (keep, archive, or drop a dead rejection). Also runs as a side-effect sub-skill when an interview or design skill calls for glossary and ADR upkeep.
Build a dependency graph over a batch of work items — issues, tasks, tickets — and surface the frontier of items that can start now. Use when planning parallel or ordered execution, checking whether two items would collide, or when another skill needs to sequence a batch.
Design and audit desktop OS integration for macOS and Windows apps, utilities, SaaS companions, developer tools, launchers, and games. Use when planning menu bar/tray behavior, installers, auto-update, startup/login items, permissions, file associations, notifications, local storage, keychain/credential storage, backup, uninstall, and OS-specific settings.
Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). Use when pretraining Llama 3.1, DeepSeek V3, or custom models at scale from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing.
Builds and optimizes language model programs with DSPy (Stanford NLP), using Signatures, modules (Predict, ChainOfThought, ReAct, ProgramOfThought) and optimizers (BootstrapFewShot, MIPRO, BootstrapFinetune). Use when replacing hand-written prompts with declarative signatures, when automatically tuning prompts against training data and a metric, when building multi-stage RAG pipelines or ReAct agents, when building classifiers or structured-output programs, or when saving and evaluating optimized LM programs across providers such as Anthropic, OpenAI, or Ollama. Not for quick prototypes or simple chains; use manual prompting or LangChain instead.
Troubleshoot Electric sync issues. Covers fast-loop detection from CDN/proxy cache key misconfiguration, stale cache diagnosis (StaleCacheError), MissingHeadersError from CORS misconfiguration, 409 shape expired handling, SSE proxy buffering (nginx proxy_buffering off, Caddy flush_interval -1), HTTP/1.1 6-connection limit in local dev (Caddy HTTP/2 proxy), WAL growth from replication slots (max_slot_wal_keep_size), Vercel CDN cache issues, and onError/backoff behavior. Load when shapes are not receiving updates, sync is slow, or errors appear in the console.
Set up a server-side proxy to forward Electric shape requests securely. Covers ELECTRIC_PROTOCOL_QUERY_PARAMS forwarding, server-side shape definition (table, where, params), content-encoding/content-length header cleanup, CORS configuration for electric-offset/electric-handle/electric-schema/electric-cursor headers, auth token injection, Bun fetch concurrency cap (BUN_CONFIG_MAX_HTTP_REQUESTS default 256), ELECTRIC_SECRET/SOURCE_SECRET server-side only, tenant isolation via WHERE positional params, onError 401 token refresh, and subset security (AND semantics). Load when creating proxy routes, adding auth, or configuring CORS for Electric.
Design Postgres schema and Electric shape definitions together for a new feature. Covers single-table shape constraint, cross-table joins using multiple shapes, WHERE clause design for tenant isolation, column selection for bandwidth optimization, replica mode choice (default vs full for old_value), enum casting in WHERE clauses, and txid handshake setup with pg_current_xact_id() for optimistic writes. Load when designing database tables for use with Electric shapes.
Build, scaffold, debug, and ship desktop applications with Electrobun — an ultra-fast TypeScript desktop app framework using Bun as the runtime and native OS webviews. This skill provides comprehensive guidance for building cross-platform desktop applications with TypeScript, covering project scaffolding, window management, main↔webview RPC communication, native UI integration, auto-updates, and distribution. Use this skill when the user mentions Electrobun, wants to build a desktop app with TypeScript or Bun, asks about BrowserWindow, BrowserView, RPC between main and webview processes, app bundling, cross-platform desktop development, or needs help with menus, tray icons, updater setup, or packaging for distribution. Also trigger when they say "desktop app" and are using or considering Bun. Common triggers include: "create Electrobun app", "desktop application TypeScript", "Bun desktop app", "BrowserWindow", "webview RPC", "Electrobun tutorial", "desktop app updater", "cross-platform desktop", "macOS app with Bun", "Windows desktop app TypeScript", or any discussion about building native desktop applications with web technologies and Bun runtime.
Build, review, debug, test, migrate, secure, or deploy Elysia and ElysiaJS applications with current documentation. Use for Bun or Node Elysia servers, routes, schemas and validation, lifecycle hooks, plugins and scope, guards, macros, context extension, error handling, Eden Treaty, OpenAPI, WebSocket, SSE, runtime adapters, security advisories, production readiness, and Elysia 1.4 to Elysia 2 beta (elysia@next, @elysia/codemod) migration. Reject route factories that pass db, Effect runtime, or other app singletons as arguments or via decorate; import those modules directly. Endpoint tests must use @elysia/eden treaty(app) — ban app.handle(new Request(...)) helpers.
Use when the user needs to write, review, or migrate SQL queries in ADVPL/TLPP using BeginSQL/EndSQL blocks and macro expressions (%table%, %xfilial%, %notDel%, %exp%, %Order%) instead of raw TCQuery string concatenation -- covering INNER/LEFT JOIN patterns, aggregations (SUM/COUNT/GROUP BY/HAVING), subqueries, column type declarations, DML via TCSqlExec, and SQL-injection prevention. Also triggers on Portuguese phrasing like "consulta SQL", "query embutida", "migrar TCQuery para BeginSQL", "escrever SELECT com filial e notDel", or "montar JOIN em ADVPL".
Generates the "what's new" release summary and diff file for changes in the Flutter engine (//engine/src/flutter) between two releases (e.g., 3.47 vs 3.44). When to use: Only activate when explicitly asked by the user to generate what's new in the engine, diff Flutter engine releases, or produce an engine release summary. When not to use: Do not use for general questions about commits, individual file history, or codebase searches unless explicitly requested to generate the engine release diff or summary.
Add Sentry v8 error tracking and performance monitoring to your services. Use this skill when adding error handling, creating new controllers, instrumenting cron jobs, or tracking database performance. ALL ERRORS MUST BE CAPTURED TO SENTRY - no exceptions.
Improve AI application with evaluation-driven development. Define eval criteria, instrument the application, build golden datasets, observe and evaluate application runs, analyze results, and produce a concrete action plan for improvements. ALWAYS USE THIS SKILL when the user asks to set up QA, add tests, add evals, evaluate, benchmark, fix wrong behaviors, improve quality, or do quality assurance for any Python project that calls an LLM model.
This skill should be used when the user asks to "execute this blueprint", "run the plan", "execute the plan", "start building from the plan", "implement the blueprint", "implement the plan", "continue the plan", "resume execution", "execute 01_milestone_name.md", "execute step 3", "run steps 3-5", or "skip to step 4".
Execute a written implementation plan with review checkpoints and structured progress tracking. Use when you have a spec or plan ready to implement, when user says "execute this plan", "implement this spec", "carry out these steps", or "start working on the plan".
Tracks ML experiments with SwanLab, an open-source tool covering swanlab.init, config and metric logging, scalar charts, and media logging (images, audio, text, GIFs, point clouds, molecules), in cloud, local (mode="local" with swanlab[dashboard]), or self-hosted setups. Includes integrations for PyTorch, Transformers, PyTorch Lightning, and Fastai. Use when logging metrics and hyperparameters for training runs, comparing runs across seeds, checkpoints, or hyperparameters, running an offline or self-hosted dashboard instead of managed SaaS, logging media alongside scalars, or adding tracking callbacks to a Transformers, Lightning, or Fastai training loop. Not for other trackers such as Weights & Biases or MLflow.
Use when the user asks to list, show, or review open issues — from the tracker recorded in docs/agents/issue-tracker.md, whether GitHub, Jira, local markdown, or another backend, or from a repository or project the user names. Read-only: not for creating, editing, closing, or triaging issues.
Orchestrate the full FinalRun loop — use finalrun-generate-test to explore a feature and author tests, run them via finalrun-use-cli, then triage failures from CLI artifacts and fix app code or tests until the run is green. Trigger for requests like "debug this failing FinalRun test", "verify and fix the feature with FinalRun", or as the Definition-of-Done loop after a UI change.
Fine-tune and serve Physical Intelligence OpenPI models (pi0, pi0-fast, pi0.5) using JAX or PyTorch backends for robot policy inference across ALOHA, DROID, and LIBERO environments. Use when adapting pi0 models to custom datasets, converting JAX checkpoints to PyTorch, running policy inference servers, or debugging norm stats and GPU memory issues
Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training. Use when need RLHF, align model with preferences, or train from human feedback. Works with HuggingFace Transformers.
Use when writing, reviewing, or refactoring TypeScript frontend code, including React components, Astro or Next.js apps, shared UI packages, client state, server-state fetching, forms, styling systems, routing, analytics, frontend tests, and frontend architecture decisions.
Implement or extend a user-facing workflow in a web application, integrating with existing backend APIs. Use when the feature is primarily a UI/UX change backed by existing APIs, affects only the web frontend, and requires following design system, routing, and testing conventions.
A global integrated development and collaboration workflow skill covering the full chain of requirement assessment, development, testing, quality, documentation, commit and release, able to integrate all basic atomic skills and automate the PDTFC+ loop with optimized division of labour and collaboration.
Comprehensive GameMaker Studio 2 and GML expertise covering beginner to advanced topics. Provides guidance on object creation, GML syntax, event systems, data structures, file I/O, shaders, networking, platform-specific features, performance optimization, and debugging. Automatically invoked when working with .gml files, .yy files, .yyp projects, or GameMaker-related tasks. References official manual (manual.gamemaker.io) for authoritative documentation.
Build visually stunning geospatial demo webapps with MapLibre globe, DuckDB-WASM, and a warm dark HUD aesthetic. Use when building satellite imagery retrieval demos, embedding visualization tools, geospatial search UIs, or any map-centric data exploration frontend. Triggers on requests to scaffold a new geo demo, build a map-based UI, or create a retrieval/search visualization over satellite data.
A Git workflow safety assistant. For branch and worktree management, protecting the main directory from being switched, skill sparse checkout, PR creation/review/merge, and commit identity and privacy checks; also supports CI quota governance, history removal of sensitive files and repository slimming. Do not use it for bulk commit message generation, project task assignment or local Agent session orchestration.
Design and refactor Go service architecture for DDD-style systems. Use when a Go codebase needs explicit ports/app/domain/adapters boundaries, dependency rules, interface placement, composition-root dependency injection, separate transport and database models, or help deciding whether DDD, Clean Architecture, or CQRS are justified.
Build RAG systems, semantic search, and document clustering with Gemini embeddings API (gemini-embedding-001). Generate 768-3072 dimension embeddings for vector search, integrate with Cloudflare Vectorize, and use 8 task types (RETRIEVAL_QUERY, RETRIEVAL_DOCUMENT, SEMANTIC_SIMILARITY) for optimized retrieval. Use when: implementing vector search with Google embeddings, building retrieval-augmented generation systems, creating semantic search features, clustering documents by meaning, integrating embeddings with Cloudflare Vectorize, optimizing dimension sizes (128-3072), or troubleshooting dimension mismatch errors, incorrect task type selections, rate limit issues (100 RPM free tier), vector normalization mistakes, or text truncation errors (2,048 token limit).
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.