Top Agent Skills

Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods. Use when fine-tuning large models (7B-70B) with limited GPU memory, when you need to train <1% of parameters with minimal accuracy loss, or for multi-adapter serving. HuggingFace's official library integrated with transformers ecosystem.

Build, test, inspect, install, and deploy MCP servers with FastMCP in Python. Use when creating a new MCP server, wrapping an API or database as MCP tools, exposing resources or prompts, or preparing a FastMCP server for Claude Code, Cursor, or HTTP deployment.

Facebook's library for efficient similarity search and clustering of dense vectors. Supports billions of vectors, GPU acceleration, and various index types (Flat, IVF, HNSW). Use for fast k-NN search, large-scale vector retrieval, or when you need pure similarity search without metadata. Best for high-performance applications.

Large Language and Vision Assistant. Enables visual instruction tuning and image-based conversations. Combines CLIP vision encoder with Vicuna/LLaMA language models. Supports multi-turn image chat, visual question answering, and instruction following. Use for vision-language chatbots or image understanding tasks. Best for conversational image analysis.

Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework. Use when training GLM models, implementing custom data generation workflows, or needing tight Megatron-LM integration for RL scaling.

Knowledge comics (知识漫画): educational, biography, tutorial.

Pitfalls, examples, and edge cases for Hermes Kanban workers. The lifecycle itself is auto-injected into every worker's system prompt as KANBAN_GUIDANCE (from agent/prompt_builder.py); this skill is what you load when you want deeper detail on specific scenarios.

Automatically collects dialectic memory to fine-tune local models.

Monitor blogs and RSS/Atom feeds via blogwatcher-cli tool.

Open-source embedding database for AI applications. Store embeddings and metadata, perform vector and full-text search, filter by metadata. Simple 4-function API. Scales from notebooks to production clusters. Use for semantic search, RAG applications, or document retrieval. Best for local development and open-source projects.

High-performance vector similarity search engine for RAG and semantic search. Use when building production RAG systems requiring fast nearest neighbor search, hybrid search with filtering, or scalable vector storage with Rust-powered performance.

Pre-built cron job templates for common automation patterns.

Play Pokemon via headless emulator + RAM reads.

Airtable REST API via curl. Records CRUD, filters, upserts.

Edit PDF text/typos/titles via nano-pdf CLI (NL prompts).

Monitor RSS feeds, arXiv, HackerNews, GitHub trending for topics of interest on cron schedule.

Send third-party WhatsApp messages or sync/search WhatsApp history via wacli, not normal active chats.

Generate charts and visualizations from structured data using D3.js and Chart.js.

Manage Trello boards, lists, and cards via the Trello REST API.

Provides guidance for training and analyzing Sparse Autoencoders (SAEs) using SAELens to decompose neural network activations into interpretable features. Use when discovering interpretable features, analyzing superposition, or studying monosemantic representations in language models.

OpenAI's general-purpose speech recognition model. Supports 99 languages, transcription, translation to English, and language identification. Six model sizes from tiny (39M params) to large (1550M params). Use for speech-to-text, podcast transcription, or multilingual audio processing. Best for robust, multilingual ASR.

Summarize or transcribe URLs, YouTube/videos, podcasts, articles, transcripts, PDFs, and local files.

Oracle CLI second-model review/debug/refactor/design with selected files, dry-run token checks, API or browser engine.

Current weather and forecasts with wttr.in via curl for locations, rain, temperature, travel planning.

Read-only EVM client: wallets, tokens, gas across 8 chains.

阶段① 需求确认。以提问方式逐步确认需求，产出 design.md 第一章。

阶段⑤ Review。审查已实现的代码并修复问题。

夜间模式：无人值守连续推进，遇歧义不停下、用最稳妥假设继续并记录。只能由用户手动开启。

阶段③ 任务划分。把 design.md 拆成 tasks/<模块>.md 和 tasks/progress.md。

Use when code is hard to read, has duplication, is too complex, has grown beyond 300 lines, or needs restructuring without changing behavior. For extracting functions, simplifying nested logic, removing dead code, improving naming, splitting large files, and applying clean code principles. Activate when user says "clean this up", "this is messy", "simplify", "extract", or "improve readability".

Use when an existing Antigravity skill has weak triggers, vague instructions, missing constraints, or poor structure and needs to be improved. Trigger on requests to rewrite, tighten, split, or review a skill for better routing and reuse.

Use this when reviewing a full system design, architecture plan, or cross-service implementation for operational risk, boundary clarity, security posture, deployability, and maintainability.

Use when user flows already exist and the team needs a screen inventory, screen responsibilities, key states, and view hierarchy for a product or feature. Trigger on requests to break a flow into pages, views, or screen-level structure.

Use this when the deliverable is a CLI tool, command suite, operator utility, or local developer tool and the team needs command structure, UX, config, packaging, and safety planning across languages.

Use this when reviewing a feature, service, API, auth design, integration, job system, or deployment plan for security weaknesses, unsafe trust assumptions, input risks, or privilege issues.

Use when implementing login, registration, session handling, protected routes, or role-based access.

Use when a product needs third-party integrations such as payments, email, analytics, AI providers, storage, or internal services. Trigger on requests to plan integration boundaries, provider selection, failure handling, or sync strategies.

Use this when a feature or product needs a clear end-to-end user journey across entry, action, system response, and outcome. Trigger on requests to map the full experience before screens or components are finalized.

Use this when the team needs detailed interaction behavior for user inputs, transitions, feedback, validation, confirmations, or stateful UI behavior that should be implementation-ready rather than purely visual.

Use this when designing or refining a Node.js backend service, daemon, worker, integration service, or internal runtime that is not just an Express route layer.

Use this when the system needs logging, metrics, tracing, alerting, or operator-visible runtime signals for APIs, workers, integrations, serverless functions, or local-to-production supportability.

Use when defining a feature's purpose and user value before building it, writing user stories, mapping user flows end-to-end, prioritizing what to build next, evaluating whether a feature solves a real problem, or reviewing scope to prevent over-engineering. Activate when user says "what should I build", "plan the feature", "is this the right approach", "define the requirements", or presents an idea without a clear user need.

Use this when the team needs to define permissions, roles, membership rules, resource ownership, or action-level authorization separate from authentication choice. Trigger on requests about access models, RBAC, ABAC, tenancy, or protected operations.

Use this when reviewing or designing UI for keyboard access, semantic structure, focus handling, labels, contrast, error communication, motion sensitivity, or screen-reader-friendly behavior.

Use this when a codebase or design effort needs consistency enforcement for tokens, spacing, typography, color usage, component variants, and reusable interaction patterns.

Use when the team needs authentication, identity providers, custom claims, session strategy, or permission design for a Firebase-backed app. Trigger on requests to design or review Firebase Auth architecture.

Use this when a Python worker, consumer, scheduled processor, or background execution service needs design or review. Trigger on requests about Python async jobs, consumers, schedulers, or long-running background processes.

Use before any complex or ambiguous task where the path forward is unclear. For breaking down multi-step problems, choosing between approaches, clarifying vague requirements, identifying hidden assumptions, planning before building, and preventing costly mistakes from starting in the wrong direction. Activate when user says "how should I approach this", "what's the best way to", "I'm not sure how to", or presents a problem without a clear solution.

Use when the team needs a strategy for validation errors, API failures, job retries, user-facing error UX, observability, or recovery behavior. Trigger on requests to design error handling across frontend, backend, and integrations.

Use this when building or planning a Python command-line tool, command surface, argument model, package entrypoint, or automation-oriented CLI workflow.

Use this when the system needs scheduled tasks, cron-style jobs, periodic maintenance, sync loops, cleanup tasks, or non-interactive background automation that is not simply an API request or queue worker.

Design an application database schema from product flows and backend responsibilities. Use when the team needs core entities, relationships, lifecycle fields, indexes, audit needs, tenancy boundaries, and migration-minded schema planning for an MVP or production system.

Use when designing PostgreSQL tables, enums, relationships, migrations, storage metadata, or tenant-aware schema structures for a Supabase-backed app. Trigger on requests about Supabase schema planning or SQL-first data design.

Use this when reviewing how a codebase organizes configuration files, env vars, runtime options, feature flags, and per-environment settings. Trigger on requests to clean up or validate config structure.

Use when building or planning Firebase Cloud Functions for webhooks, backend APIs, scheduled jobs, claim assignment, or privileged Firestore logic. Trigger on requests to create or review Firebase Functions.

Use when a web product or component needs responsive layout planning across mobile, tablet, and desktop. Trigger on requests about breakpoints, layout adaptation, content priority, or mobile-first UI behavior.

Use this when the team needs to decide how responsibilities should be split across services, modules, workers, APIs, or runtimes. Trigger on requests about service boundaries, ownership, modularity, monolith versus separation, or where logic should live.

GSD Execute — Get Shit Done

Use when the app is slow, pages take too long to load, API responses exceed 200ms, memory usage is high, or renders are excessive. For profiling, identifying bottlenecks, optimizing React renders, reducing bundle size, improving database query speed, and measuring before/after improvements. Activate when user says "it's slow", "optimize this", "reduce load time", or "too many re-renders".

Use this when designing scripts or small tools for automation, maintenance, CI tasks, file manipulation, config generation, or local workflow support across shell, Python, or mixed environments.