AI Agent Engineer

AI Agent Engineer
PERSONAL PROGRESS

Mark this course

Course content is public. Sign in to save personal progress.

Agent Skills Learning Path

6 min read

Agent Skills

Course overview

Agent Skills are an open format for giving an AI agent reusable specialist capabilities through a folder containing a SKILL.md file, scripts, and supporting resources. They package the procedural knowledge needed for a recurring class of work into something discoverable, selectively loaded, versioned, and testable. This course first establishes the boundaries between the format, clients, and authorization; it then covers triggering, progressive disclosure, safety, and evaluation. The eight pre-existing official materials remain available as an upstream-reference layer.

Source and scope

The official documentation index is agentskills.io/llms.txt; the specification and authoring guidance were checked on 2026-07-22. name and description remain required. license, compatibility, and metadata are optional, while allowed-tools is still explicitly experimental. Install locations, trigger behavior, tool authorization, and confirmation prompts remain client-specific. Do not mistake one product’s behavior for a universal guarantee. In particular, the presence or loading of a Skill, or the allowed-tools field, does not constitute user authorization, identity verification, or evidence that a script is safe.

The licensing boundary was rechecked on 2026-07-20: the upstream repository code uses Apache-2.0, while its docs/ directory and website documentation use CC BY 4.0. They are not interchangeable. Before public release, every upstream-reference page must retain Agent Skills attribution, its original link, and a notice of Chinese translation, editorial organization, and formatting changes. If those fields are absent, publish only a source-link page. An example Skill that independently declares CC0-1.0 follows its own declaration.

Where this course fits

This course belongs to the “Single-Agent Systems and Tools” stage. After learning the task loop in Agent Core and the action contract in Tool Calling, you can package stable workflows as Skills. When external tools or data are involved, pair them with MCP. A Skill does not replace Tool Calling, MCP, or an authorization system: it supplies workflow context, while effective permission is enforced by the host, tool service, and business system. This is a capability dependency after choosing this branch, not a requirement that every role finish every linked course; if you have no reusable workflow, you can defer Agent Skills.

Learning objectives

  • Explain the boundary between the open format and a particular agent client.
  • Distinguish Skills, persistent instructions, Tool Calling, MCP, and effective authorization.
  • Create a specification-aligned folder and SKILL.md YAML frontmatter.
  • Control context cost with the metadata → instructions → resources layers of progressive disclosure.
  • Write a description that covers genuine intent without over-triggering.
  • Design non-interactive, safe-by-default scripts with structured output, and organize references and assets only when needed.
  • Establish source, version, diff-review, and isolated-trial boundaries for third-party Skills and updates.
  • Improve quality with format validation, trigger tests, with/without-Skill baselines, and observable assertions.

Prerequisites

  • Read and write Markdown and simple YAML frontmatter.
  • Run Python 3 scripts in PowerShell 7 and inspect exit codes.
  • Understand an agent context window, Tool Calling, and least privilege.
  • No prior knowledge of a specific client is required; read that client’s current official documentation before installing anything.
  1. Positioning, Client Differences, and Permission Boundaries — establish the responsibility boundary between the format, tool protocols, and authorization.
  2. Open Format and Directory Structure — separate the specification core from each client’s implementation choices.
  3. SKILL.md Frontmatter and Progressive Disclosure — write accurate frontmatter and split context on demand.
  4. Trigger Descriptions and Scope Boundaries — calibrate a description with realistic positive and negative cases.
  5. Scripts, Resources, and Safety — design executable resources, supply-chain review, and least-privilege boundaries.
  6. Testing, Evaluation, and Iteration — compare against a baseline and improve from evidence.
  7. Project: Create and Validate an Agent Skill — run a complete, offline example.

Hands-on entry points

  • Main project: Create and validate the text-statistics Skill.
  • Trigger exercise: write eight to ten realistic positive requests and the same number of negative requests for your own Skill. Prefer nearby tasks that are easy to trigger by mistake, then reserve a separate validation set.
  • Evaluation exercise: write an assertion for one mechanically verifiable outcome, and compare it with the baseline where the Skill is not loaded.

Mastery checklist

  • I can create a valid folder with SKILL.md from scratch and explain why name must match its parent directory.
  • I can explain how a target client discovers, selects, or preloads Skills without treating progressive disclosure as a guarantee for every host.
  • I can write a description that says both what the Skill does and when to use it, then test its boundary with near-miss negative cases.
  • I can review a script’s network, file, command, credential, and output boundaries, and review a third-party Skill’s source, revision, and update diff.
  • I can explain why a Skill, allowed-tools, and an MCP resource do not prove user intent or authorization.
  • I can run the local example validator and scripts, and understand both their coverage and their limits.
  • I can use a with-Skill/without-Skill baseline, concrete assertions, and human review to decide whether to revise a Skill.

Relationships to other courses

  • Prompt Engineering focuses on one-off or templated instructions. A Skill packages a stable workflow, scripts, and reference material, and is selected through its description.
  • Context Engineering explains why information should load only when needed; progressive disclosure is a concrete engineering application.
  • Tool Calling defines the parameter, result, and error contract for one action. A Skill provides the workflow context for when to call it, how to verify it, and when to stop.
  • MCP exposes tools, resources, and prompts. A Skill can teach an agent when and how to use them safely, but grants no access and does not replace server-side validation.
  • Evaluation Framework supplies broader methods for experimental design, metrics, and error analysis once its full course is available.

Primary references

All sources below are first-party materials, retrieved or checked on 2026-07-22:

Official-documentation reference layer

The files below are pre-existing Chinese editorial versions of official documentation. Their body text, provenance data, and layer status are intentionally retained. Their client lists and product-specific operations represent a captured snapshot only: current installation instructions, commands, permissions, and support must be checked against the target client’s official documentation and local verification. The learning route above adds the operational boundary around that reference layer.

01 — Overview

02 — Skill authors

Backlinks