发现 Skills

Pythonic idioms, PEP 8 standards, type hints, and best practices for building robust, efficient, and maintainable Python applications.

For cross-cutting concerns: add behavior without modifying functions, caching, timing, logging, validation wrappers.

For persistent state: closures capture outer variables, alternative to classes for simple state, factory functions that remember context.

For flexible parsing: attempt multiple type conversions in a defined order, gracefully falling back from specific to more general types.

For 2D debugging: visualize grid/board state, show puzzle progress, make algorithm behavior visible.

For computational geometry: convex hull, point enclosure, polygon operations. Uses the Monotone Chain (Andrew) algorithm with stack-based turn (cross-product) detection.

Manage local Ollama LLM models for development and testing. Use when: running local models, configuring Ollama, switching between fast/quality models, optimizing VRAM usage, troubleshooting model performance, creating Modelfiles, or integrating local LLMs with applications via OpenAI-compatible API (http://localhost:11434/v1). Triggers: ollama, local model, local LLM, VRAM, GPU memory, model speed, inference, Modelfile, llama.cpp, quantization.

For generic algorithms: strategy pattern, callbacks, configurable behavior. Pass functions as parameters to customize algorithm behavior.

For performance reporting: timing wrappers, throughput calculations, profiling summaries.

For data transformation: list/dict/set comprehensions, chained filtering and mapping, creating lookup structures concisely.

For solution checking: independent verification that result is correct, separate from how it was computed.

For writing clean functions: start with a docstring, iterate until the code matches the docstring exactly, achieving docstring-code isomorphism.

Evaluate Agent Skill design quality against official specifications and best practices. Use when reviewing, auditing, or improving SKILL.md files and skill packages. Provides multi-dimensional scoring and actionable improvement suggestions.

Use when writing prose humans will read—documentation, commit messages, error messages, explanations, reports, or UI text. Applies Strunk's timeless rules for clearer, stronger, more professional writing.

Songwriting craft, AI music generation prompts (Suno focus), parody/adaptation techniques, phonetic tricks, and lessons learned. These are tools and ideas, not rules. Break any of them when the art calls for it.

Fine-tune large language models (LLMs) using reinforcement learning with TRL — SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and tools for training reward models. Use when you need RLHF, want to align a model with human preferences, or train from human feedback. Works with Hugging Face Transformers.

Request movies and TV shows via Overseerr, monitor request status, and manage media requests. Use when the user asks to "request a movie", "request a TV show", "check request status", "pending requests", "Overseerr status", "media request", or mentions Overseerr/Seerr media requesting.

Modbus industrial protocol implementation in Rust, Go, and Python. Trigger: modbus, plc, scada, industrial protocol, registers, coils

Generate a comprehensive monthly summary — income/spend/savings rate, budget vs actual, overrun categories, top categories and payees, largest transactions, net worth at month end. Use when the user asks for "monthly report", "month recap", "how did the month go", "summary of March".

Analyze the current project context and existing `.agents/skills` to design and create high-value subagents. Use this when the user asks to create subagents, specialize agent behavior, or convert repeated skill workflows into isolated, reusable agents. Always read `AGENTS.md`, `.agents/rules`, `.agents/skills`, `.agents/agents`, and `.agents/ai-scaffolding-context.md` first, when available, before proposing or creating subagents.

Standardize OpenAI API direct-model integration for chat, vision, image generation, structured outputs, reasoning effort, streaming, and tool calling. Use when Codex needs to choose an OpenAI model, wire Responses API or Chat Completions compatible request/response flows, expose reasoning effort, JSON object, JSON schema, tool/function calling, stream controls, model selectors, or shared progress and error states. Do not use for ChatGPT consumer-app settings, Apps SDK orchestration, Agents SDK orchestration, non-OpenAI providers, or prompt-only work with no integration change.

Inter-agent communication for multi-agent workflows. Use when multiple agents need to coordinate, share information, or reserve resources.

Structural audit for AI writing patterns. Catches paragraph-level tells that word-level checks miss: false agency, binary contrasts, dramatic fragmentation, narrator-from-a-distance, formulaic rhythm. Scores on 5 dimensions (35/50 threshold). Use on content pipeline output before publishing. Invoke with /stop-slop.

Use when smoke testing or verifying the web app on staging (staging.example.com) or production (app.example.com). Covers the full user flow without DevToolbar, admin login, and deployment verification.