"Compare these tools and write a report with sources." This task needs clear comparison criteria, evidence standards, and an output format. It also requires documentation, release notes, and user feedback.
Installing a research Skill lets an assistant reuse a research process. Connecting search or repository tools lets it obtain the material. Together, they form a workflow the assistant can actually carry out.
A Skill Captures a Method for the Task
Agent Skills is an open format for packaging specialized knowledge and workflows. A skill directory contains an entry-point Markdown file and may include references, templates, or scripts. Scripts are optional in the official specification: a set of text-only review rules can also be a skill.
For example, a research-report Skill might require the assistant to define the products and versions being compared, prioritize official sources, separate facts from opinions and assumptions, retain sources for important judgments, and use consistent criteria in the final report. These requirements are reusable; they do not need to be assembled from scratch for every task.
A simplified directory might look like this:
research-brief/
SKILL.md
references/
assets/
The entry description tells the assistant when to use the skill. The body explains the method, while the supporting directories hold material to read when needed. The official Agent Skills introduction describes progressive disclosure: discover the name and description first, read the full instructions when the task matches, and then load resources or execute code as needed.
The client still decides how skills are invoked. Some can be matched automatically; others require manual selection. Supporting the format does not mean the client has every tool mentioned in a skill's instructions.
MCP Provides an Interface to Tools and Services
A research method does not produce search results on its own. An assistant still needs tool connections to inspect a repository, read online documentation, or access business data.
MCP defines how AI applications interact with tools and data services. GitHub's official MCP Server, for example, can provide capabilities such as reading files, querying Issues, and inspecting Pull Requests. The assistant obtains material through these tools and processes it according to the research Skill.
There is also some overlap. MCP can provide Prompts, and a Skill can include scripts and instructions for using tools. It is therefore misleading to reduce the distinction to "one only writes prompts, the other only performs actions." A more useful distinction is that a Skill packages a reusable method, while MCP provides a communication interface for accessing services.
What Each Contributes to the Same Report
Assuming the assistant already has the tools it needs, a report task might proceed as follows:
| Stage | What the Skill Contributes | What the Tool Connections Contribute |
|---|---|---|
| Define the scope | Set comparison criteria, versions, and evidence standards | Retrieve relevant project information |
| Gather material | Specify source priorities and recording practices | Read documentation, search Issues, or retrieve data |
| Reach conclusions | Separate facts from assumptions and organize the comparison | Investigate missing evidence |
| Deliver the report | Check sources, formatting, and omissions | Output or save the result as needed |
This is a workflow design, not a guarantee of success. If documentation cannot be retrieved, the assistant should identify the gap. If a tool returns outdated material, the research rules should require a version check. The instructions and the tool results have different responsibilities; neither substitutes for the other.
How to Identify the Missing Part
If the assistant can obtain material but repeatedly misses comparison criteria, improve the Skill's instructions first. If the method is clear but the required data is inaccessible, add or repair the tool connection. If the output is useful but inconsistent in format, templates and delivery checks may help more.
Evaluate the combination by looking at its output: does the report cover the agreed products, provide sources for important judgments, and use material that applies to the target versions? The number of installed skills and tools cannot answer those questions.
For a specific task, connecting the method, material retrieval, and delivery checks makes effectiveness easier to judge than adding more names to a list.