superpowers-zh

Implements a Model Context Protocol (MCP) server that registers discrete coding skills as callable tools, enabling Claude Code, Copilot CLI, Cursor, Windsurf, and 11+ other AI coding agents to discover and invoke skills through a standardized schema-based function registry. Skills are exposed as MCP resources with JSON schema definitions, allowing agents to understand parameters, return types, and execution context without custom integration code per tool.

Bundles 6 original Chinese-language coding skills (test generation, code review, refactoring, documentation, debugging, architecture design) as pre-crafted prompt templates that are optimized for agentic execution. Each skill encodes best practices (TDD-first approach, structured output formats, error handling patterns) as system prompts that guide LLM behavior without requiring fine-tuning, enabling consistent, high-quality code generation across different LLM backends.

Enables versioning of skill prompts and automatic A/B testing to compare different prompt versions. Routes a percentage of skill invocations to different prompt versions and collects metrics (execution time, output quality, user satisfaction) to determine which version performs better. Automatically promotes high-performing versions and deprecates low-performing ones. Supports gradual rollout (canary deployment) to minimize risk of bad prompt changes.

Provides a framework for developers to create custom skills by defining prompt templates, input/output schemas, and execution logic. Custom skills are registered in the MCP server and exposed to all connected AI agents with the same interface as built-in skills. Includes TypeScript/JavaScript SDK with type definitions, validation helpers, and testing utilities. Supports skill packaging and distribution via npm for community sharing.

Routes skill invocations across multiple LLM providers (OpenAI, Anthropic, Google, local Ollama, Chinese providers like Qwen/Baichuan) with automatic fallback logic. Detects provider availability, handles rate limits, and retries failed requests using exponential backoff. Abstracts provider-specific API differences (function calling schemas, token counting, context window limits) behind a unified skill execution interface, enabling skills to run on any available LLM without code changes.

Automatically extracts and injects relevant codebase context (imports, type definitions, related functions, test files, documentation) into skill prompts before LLM execution. Uses AST parsing and semantic analysis to identify code dependencies and include only relevant context (not entire codebase), staying within LLM context windows. Caches parsed codebase structure to avoid re-parsing on repeated skill invocations, reducing latency by 70-80%.

Defines and enforces JSON schema constraints on skill outputs (code review comments, refactoring suggestions, test cases) to ensure machine-parseable, consistent results. Uses schema validation and retry logic — if LLM output violates schema, automatically re-prompts with schema examples and stricter instructions. Supports schema versioning to enable backward compatibility as skills evolve.

Enables sequential execution of multiple skills with automatic data flow between steps (output of one skill becomes input to next). Provides a workflow DSL (YAML or JSON) to define skill chains, with conditional branching (if code review fails, run refactoring skill), error handling (retry failed steps, skip on error), and result aggregation (combine results from parallel skill invocations). Executes chains with dependency tracking to optimize parallelization where possible.

Provides 6 original coding skills with prompts optimized for Chinese language and Chinese development practices. Includes cultural conventions (naming patterns, code organization, documentation style) common in Chinese teams, and is optimized for Chinese LLMs (Qwen, Baichuan) which may have different training data and instruction-following patterns than English-trained models. Supports bilingual prompts (Chinese + English) for mixed-language teams.

Integrates with Ollama (local LLM runtime) to execute skills entirely on-device without cloud API calls. Supports popular open-source models (Llama 2, Mistral, CodeLlama) optimized for code generation. Automatically detects Ollama availability and falls back to cloud LLMs if local model is unavailable. Caches model weights locally to avoid re-downloading, reducing startup time from minutes to seconds on subsequent runs.

Provides MCP client implementations for popular code editors (VS Code, Cursor, Windsurf, Vim, Neovim) that connect to the superpowers-zh MCP server and expose skills as editor commands and inline suggestions. Integrates with editor keybindings, context menus, and command palettes for seamless workflow. Supports editor-specific features (VS Code inline code actions, Cursor's native AI features, Vim's command mode) while maintaining consistent skill behavior across editors.

Provides comprehensive logging and monitoring of skill execution including execution time, token usage, LLM provider, error details, and result quality metrics. Outputs structured logs (JSON format) that can be ingested into observability platforms (Datadog, New Relic, ELK Stack) for real-time monitoring and alerting. Includes built-in metrics (success rate, average latency, cost per skill) and dashboards for tracking skill usage and performance over time.