Understand Skills by separating them from other mechanisms

Compare Skills with model weights, MCP, tools and retrieval.

Suggested learning time: 45 minutes.

1. 1Task request
2. 2Reusable Skill procedure
3. 3Tools or retrieved sources
4. 4Reviewed deliverable

Original learning map: arrows show the reading or decision sequence, not a measured execution trace.

Evidence and exercise status

This chapter is an editorial learning guide. Reading sources is distinct from executing a Skill and measuring its effect. Exercise status: not-run. No experiment logs or model outputs exist.

About this course

A ten-chapter course on mechanisms, how to read established examples, safe custom authoring, and reproducible comparisons, learned through one small Skill.

Audience: learners using AI for creative or development work who want to judge Skill behavior themselves. Editorial version: 2026.10.03-v1. Source verification date: 2026-10-03. Evidence status: editorial-guide-with-source-reviewed-catalog.

Learning goals

• Explain the differences between a Skill, MCP, a model, and search.
• Distinguish problems that need a Skill from those that do not.

Roles in the work

• Learner: define hypotheses and grading criteria.
• AI: assist with reading and deliverable creation within the authorized scope.
• Reviewer: inspect outcomes and logs separately.

Inputs

• The input examples specified in this chapter.
• The official material and versions to verify.

A Skill carries reusable knowledge about work

At the center of Agent Skills is a folder containing SKILL.md. It holds a description of its purpose and instructions for the work, with scripts, references, or templates when needed. It does not replace model weights. For example, an editorial rule that every textbook chapter must include goals, an exercise, and acceptance criteria can be reused without rewriting it into a long request each time. When output changes after a Skill is read, the runtime context and procedure have changed. That operation alone does not mean the model has permanently retrained.

Agent Skills overview

Separate connections, instructions, and knowledge

MCP is a standard connecting AI applications with external systems, providing tools and information sources. A Skill can explain when to use them, in what order, and against which criteria. The means of acting, like a kitchen knife, and knowledge of the procedure, like a recipe, are different. In practice they can overlap: a Skill can bundle a script or retrieve references through MCP. Think in two layers—connected capabilities and knowledge for using them—instead of choosing between MCP and Skills.

MCP introduction

Six layers that are often confused

Model weights are learned numbers; system instructions constrain the whole run; the user request states the current purpose; search and RAG retrieve needed information; tools perform actual operations; and a workflow designs the process. A Skill packages procedures and resources. It can express a workflow, but reading it does not turn it into a deterministic program that always executes every step. Enforce essential access controls and spending limits in the execution environment as well as in written instructions. This is a design distinction used throughout this course.

Choose one question before adopting a Skill

“Make the AI smarter” is not a measurable effect. “Reduce how often sources and conjecture are mixed in a Japanese report” is measurable. First record one existing failure, the input conditions that produce it, and the desired deliverable. A Skill may be a candidate when you need a consistent format, a specific glossary, or a verification step that must not be skipped. If a short one-time request is enough, or the input material itself is missing, adding Skills will not solve that underlying gap.

Terms to start with

Term	Meaning
Skill	A reusable unit combining a purpose description, work instructions, and required resources.
Progressive disclosure	A design that reads candidate descriptions, the body, and necessary resources in stages.
Harness / runtime	The application-side environment providing a model with context, tools, permissions, and execution flow.
Baseline	The reference condition for a comparison; specify whether assistance other than the target Skill is present too.
Contamination	Output, instructions, generated files, or other material from one comparison condition entering the other.
Ablation	A comparison that removes a component to investigate its contribution.

Workflow

1. Separate “read a PDF textbook, organize its main points, and publish them on the web” into the six layers.
2. Decide whether the Skill body alone can authorize publication.
3. Describe one failure you want the Skill to reduce in observable terms.

Outputs

• A six-layer mapping note and one measurable hypothesis.

Quality checklist

• You do not confuse SKILL.md with model weights.
• You do not describe MCP as merely a prompt collection.
• You have named one failure to improve.

Failure diagnosis

• Symptom: Recording success without observing the effect.
• Cause: Confusing expected judgments with actual outputs.
• Fix: Keep unexecuted work as not-run, clear measurement fields, and obtain raw outputs and logs before scoring.

Exercise: Organize a task into six layers

Follow the workflow above in order and create the stated deliverable.

Completion criteria: Separate authorization for external operations from work procedures, and do not count permanent model learning as a Skill effect.

Status: not-run.

Source scope

Sources support feature descriptions and distributor statements in the text and catalog. They are not evidence of measured effects or popularity ranks. Verification dates record reading public sources, rather than publication or update dates. Rolling references such as main are not pinned experimental versions.

Offline experiment worksheet

Current Learn edition

The chapter set was restored to the common Learn archive on 2026-10-04. Original source reviews retain their 2026-10-03 date. The former shared experiment screen is not available in this edition: use the linked offline worksheet and keep every planned run marked not-run until observations exist. No experiment, ranking, or performance measurement was created by this restoration.