# Agent Skills for Building AI Agents Give your coding agent reusable guidance for designing tools, writing agent instructions, and building eval cases. Each skill packages guidance from Agentailor articles into a SKILL.md file and supporting references. A compatible coding agent can load that guidance when the task matches. Choose the skill that fits your work, install it, then describe what you want to do. Use with Claude Code or a coding agent that supports the [Agent Skills specification](https://agentskills.io/specification). Source repository: https://github.com/agentailor/skills ## Tool Design Design tools an AI agent can actually use — framework- and language-agnostic. ### When to use it You are adding tools or your agent picks the wrong tool, sends invalid inputs, or gets too much output. ### What it helps you produce Clearer tool definitions, focused outputs, and a checklist to review them. ### What the skill covers Five production-tested principles for tools an agent picks correctly, calls with valid inputs, and reads back without flooding its context — plus a validation checklist and worked examples. Applies whether the tool is an MCP server tool, a LangChain/LangGraph tool, or a plain function-calling definition, in TypeScript, Python, or otherwise. ### Install Run in your project directory with Node.js and npx available. The installer lets you choose your coding agent. ```sh npx skills add agentailor/skills --skill tool-design ``` For all your projects: ```sh npx skills add agentailor/skills --skill tool-design --global ``` Or ask your coding agent in the project where you want to use the skill: Install the tool-design skill from https://github.com/agentailor/skills into this project for the coding agent I'm using. ### After installing, try this Use the tool-design skill to review my MCP tools for ambiguous names, invalid-input risks, and oversized responses. Propose concrete improvements. Sources: [repo](https://github.com/agentailor/skills/tree/main/tool-design) · [Writing Effective Tools for AI Agents](https://blog.agentailor.com/blog/writing-tools-for-ai-agents) ## Agent Prompt Engineering Design system prompts for autonomous, tool-using agents that run in a loop. ### When to use it You are writing agent instructions or your agent makes poor decisions about its next step. ### What it helps you produce Instructions with clearer decisions, boundaries, and stopping conditions. ### What the skill covers The principles, heuristics, and evaluation strategy that make an agent reliable when it decides its own steps — with worked prompt examples and the anti-patterns to avoid. Reach for it when writing agent instructions, a system prompt, or an agent config, or when an agent is making poor decisions. ### Install Run in your project directory with Node.js and npx available. The installer lets you choose your coding agent. ```sh npx skills add agentailor/skills --skill agent-prompt-engineering ``` For all your projects: ```sh npx skills add agentailor/skills --skill agent-prompt-engineering --global ``` Or ask your coding agent in the project where you want to use the skill: Install the agent-prompt-engineering skill from https://github.com/agentailor/skills into this project for the coding agent I'm using. ### After installing, try this Use the agent-prompt-engineering skill to review this agent’s instructions for unclear decisions and stopping conditions. Suggest changes and explain how to evaluate them. Sources: [repo](https://github.com/agentailor/skills/tree/main/agent-prompt-engineering) · [The Art of Agent Prompting](https://blog.agentailor.com/blog/the-art-of-agent-prompting) ## Agent Eval Cases Decide which agent behaviors are worth an eval case, then write those cases. ### When to use it You have observed agent failures and need regression cases in your existing eval harness. ### What it helps you produce Focused cases with appropriate graders, based on failures you have actually seen. ### What the skill covers The workflow that turns observed failures into a small, honest eval suite: elicit real failures instead of inventing them, push each down to the cheapest layer that can catch it, group what survives by defect class, pair every case that pushes a behavior with one that bounds it, and pick graders by whether the target has one spelling or many. Also covers reading the first red run — why a case that passes before you fixed anything is a weak test. Harness-agnostic: it writes into whatever suite you already have. ### Install Run in your project directory with Node.js and npx available. The installer lets you choose your coding agent. ```sh npx skills add agentailor/skills --skill agent-eval-cases ``` For all your projects: ```sh npx skills add agentailor/skills --skill agent-eval-cases --global ``` Or ask your coding agent in the project where you want to use the skill: Install the agent-eval-cases skill from https://github.com/agentailor/skills into this project for the coding agent I'm using. ### After installing, try this Use the agent-eval-cases skill to turn these observed failures into focused eval cases in my existing harness. Ask me for failure examples if I have not provided them. Sources: [repo](https://github.com/agentailor/skills/tree/main/agent-eval-cases) · [How to Write Your First AI Agent Evals](https://blog.agentailor.com/blog/first-eval-cases) For article search and repo docs, connect via MCP: https://agentailor.com/for-agents#mcp. Skills and MCP can be used together. --- Source: https://agentailor.com/skills