LLM 與 Agent

Enhance and refine prompts for AI coding agents using Chain-of-Thought reasoning. Use when user asks to improve, optimize, rewrite, or enhance a prompt. Transforms vague requests into structured, high-context instructions for Codex, Claude Code, or Gemini CLI.

Parse Claude command frontmatter and expose body/meta for Codex CLI.

LangGraph development for stateful multi-agent applications, cyclic workflows, conditional routing, human-in-the-loop patterns, and persistent state management. Use for complex AI orchestration, agent coordination, and production-grade agentic systems.

Configuring and optimizing 16-bit Low-Rank Adaptation (LoRA) and Rank-Stabilized LoRA (rsLoRA) for efficient LLM fine-tuning using triggers like lora, qlora, rslora, rank selection, lora_alpha, lora_dropout, and target_modules.

AI assistant framework for building unique, authentic portfolio websites from scratch. Guides agents through research, design, and implementation phases.

Conducts structured interviews to produce Change Contracts - definitions of done expressed in terms of behavior, constraints, operability, and acceptance criteria. Use when externalizing intent for work to be delegated to AI agents, when defining requirements without implementation details, or when asked to create a change contract.

Use when designing database schema, writing migrations, or making data storage decisions - enforces UUIDs, account_id multi-tenancy, state-as-records, no foreign keys, and proper index patterns

Execute agentic coding tools directly via bash commands with output capture. Use this when the user requests to run AI agents (Gemini, Claude Code, etc.) and get their results directly.

Transcribe audio via OpenAI Audio Transcriptions API (Whisper).

Use when spawning Codex agents via Otto to delegate implementation work.

Convert MCP servers into standalone Bash-invokable scripts. Use when user wants to make an MCP server usable as bash commands, convert MCP to CLI, or wrap MCP tools for agent use.

Write minimal, strategic tests focused on core user flows and critical paths during feature development, deferring comprehensive edge case testing. Use this skill when writing unit tests, integration tests, or end-to-end tests, creating test files (*.test.ts, *.spec.js, __tests__/*), testing critical user workflows, focusing on behavior rather than implementation details, using descriptive test names, mocking external dependencies (databases, APIs, file systems), ensuring tests run quickly, testing business-critical functionality, or determining which tests to write during development versus dedicated testing phases. Apply this skill when completing features at logical milestones, writing tests for primary user flows, or balancing feature development speed with adequate test coverage.

Prompt-native architecture patterns, agent orchestration, iteration loops, and state management for autonomous work.

Multi-remote git push workflow for GitHub and Azure DevOps. Use when the user says"push to github", "push to azure", "push to origin", "push to secondary", "push to devops","push to both", "sync remotes", or any variation requesting git push operations.Handles the azure-sync branch workflow that excludes .claude/, specs/, and CLAUDE.md from Azure.

Career Advisor agent for PhD application planning and career management. Helps with university research, application tracking, professor matching, and career documentation.

Comprehensive manual for Telegram Mini Apps SDK. Use when developers need guidance on creating web applications inside Telegram, working with WebApp API, managing user data, handling authentication via initData, implementing buttons and events, working with storage, and integrating with Telegram ecosystem features.

Development guide for Claude Agent SDK integration in IPA Platform.Use when:- Building autonomous AI agents with Claude- Implementing agents that need to make independent decisions- Creating agents with tool usage, MCP integration, or subagent delegation- Integrating Claude SDK with existing Microsoft Agent Framework- Working with code in backend/src/integrations/claude_sdk/Provides: query(), ClaudeSDKClient, built-in tools, hooks, MCP, subagents patterns.

This skill should be used when the user asks to "orchestrate jira", "work on issue", "complete jira ticket", "development workflow", "jira automation", "issue lifecycle", "work on story", "fix bug ticket", or needs guidance on coordinating development work through Jira with multi-agent orchestration patterns.

Systematic documentation audit that validates every documentation claim against code and identifies undocumented features - executable as a repeatable Claude Code skill (project)

Send messages and notifications to Poke (poke.com) via webhook API. Use when alerting the user, sending task completion notifications, status updates, reminders, or any message to their Poke assistant.