Skill pack — Specialist

AI Agent & MCP Server Builder

Build production MCP servers and agent workflows, then prove they work with a reusable eval harness.

What you get

agent-eval-engineer110 lines
Builds and runs evaluation suites for LLM features and agents: case sets from real traffic, deterministic and LLM-judge checks, tool-call assertions, prompt-injection suites, pass@k for flaky outputs, and CI regression gates with cost and latency tracking. Use when shipping or changing a prompt, model, agent or MCP tool, or when quality "feels worse" with no numbers.
agent-workflow-engineer123 lines
Designs reliable LLM agent workflows: picks the simplest pattern that works, designs tools, manages context, enforces step/token/time/spend budgets, adds human approval for irreversible actions, and defends against prompt injection. Use when building or reviewing an agent, multi-step LLM pipeline, tool-calling loop, or when an agent loops, overspends, or acts on untrusted content.
mcp-server-builder119 lines
Designs, builds, secures and tests production Model Context Protocol (MCP) servers in TypeScript with the official SDK. Use when creating or reviewing an MCP server, adding tools/resources/prompts, choosing stdio vs Streamable HTTP, adding OAuth to a remote server, or debugging a host that cannot see or call your tools.

Every pack ships with eval prompts that catch the mistakes generic agents make. A few of the 14 scenarios:

Checking access…