Agent Skills Turn Reusable Professional Know-How Into Structured Directories, Not Just Saved Prompts
Every time I ask AI to organize meeting minutes, I have to explain again: remove filler words, categorize by topic, list responsible persons and deadlines, and don't fabricate uncertain information.
The next time I use a different meeting record, I have to write the same requirements again. If there are five people on the team, each person might write five different versions of the prompt.
What these tasks really lack is not a longer ad-hoc prompt, but a working method that can be saved, reused, and continuously improved.
Agent Skill is designed to carry this method.
Skill does not add new parameters to the model
Anthropic's official Skills repository defines a Skill as a set of instructions, scripts, and resources dynamically loaded by an Agent to improve the execution of specific tasks.
It does not retrain the model, nor does it modify model parameters. A Skill is essentially a directory that stores the following for a certain type of task:
- Trigger conditions
- Execution steps
- Output format
- Quality standards
- Reference materials
- Executable scripts
- Templates and other resources
For example, a meeting minutes Skill can tell the Agent:
Receive meeting transcript
→ Clean filler words and irrelevant content
→ Identify meeting topics
→ Consolidate viewpoints by topic
→ Extract conclusions and action items
→ Output in a fixed format
The model inherently has the ability to summarize text; what Skill solves is "what professional process should be followed for summarization, and what result qualifies as acceptable."
A minimal Skill only has one SKILL.md
According to the Agent Skills specification, a Skill must contain at least one SKILL.md:
meeting-minutes/
└── SKILL.md
SKILL.md consists of two parts:
---
name: meeting-minutes
description: Meeting minutes generator. Use when the user provides a meeting transcript or recording transcription and needs to organize it into structured meeting minutes.
---
# Meeting Minutes Generator
Generate structured meeting minutes based on the meeting transcript provided by the user.
## Workflow
1. Read through the full text, clean irrelevant content
2. Organize meeting content by topic
3. Extract meeting conclusions and action items
4. Output according to a fixed template
The part enclosed by --- is YAML Frontmatter, and below it is the task description in Markdown format.
In the minimal structure, name and description are required fields.
name is the stable identifier of the Skill
name is used to identify the Skill. The specification requires it to use lowercase letters, numbers, and hyphens, and to be consistent with the folder name.
name: meeting-minutes
The following names are not suitable as standard Skill names:
name: Meeting Minutes
name: meeting_minutes
name: -meeting-minutes
The name is not responsible for explaining all functions; it just needs to be short, stable, and identifiable. What really determines when a Skill is used is the description.
description determines when the Skill is triggered
The Agent will not wait for the user to accurately say the Skill name before invoking it.
The user might say:
Help me organize meeting minutes
Summarize this meeting
Organize the recording transcription into meeting notes
Extract to-do items from the meeting
These expressions are different, but they point to the same type of task. The description needs to explain simultaneously:
What this Skill can do
Under what circumstances it should be used
The description used by the meeting minutes Skill is as follows:
description: Meeting minutes generator. Use when the user provides a meeting transcript, recording-to-text, or meeting notes and needs to organize them into structured meeting minutes. This skill should be used whenever the task involves converting meeting conversations into structured minutes.
If you only write:
description: Helps with meetings
It is difficult for the Agent to determine whether "schedule a meeting time," "create a meeting link," and "generate meeting minutes" should all trigger it.
If the description is written too narrowly, it might not be found when needed; if written too broadly, it will conflict with other Skills.
The body stores the execution method for professional tasks
The YAML header solves "when to use," and the Markdown body solves "what to do after triggering."
The meeting minutes Skill first requires cleaning the noise from the original transcription:
### Step 1: Read through the full text, clean irrelevant content
- Filler words and pet phrases: such as "um," "ah," "that," "and then"
- Chit-chat unrelated to the meeting
- Repeated expressions of the same viewpoint
- Technical interruptions like "can you hear me," "bad signal"
Keep all substantive discussions, viewpoints, decisions, and action items.
Then it specifies the output structure:
## Meeting Minutes
### 1. Basic Meeting Information
### 2. Meeting Objectives
### 3. Meeting Content
### 4. Action Items
Finally, it sets quality boundaries:
1. Leave uncertain information blank, do not fabricate
2. Similar discussions scattered at different time points should be grouped under the same topic
3. Main viewpoints should be attributed to speakers as much as possible
4. Action items include responsible person, task description, and deadline
5. Use concise bullet points and tables, do not write long, rambling accounts
What is saved here is not just an output template, but also the judgment rules used by professionals when handling meeting materials.
Why doesn't the Agent read all Skills at once?
An Agent might have dozens or even hundreds of Skills installed simultaneously. If all instructions, scripts, and reference materials were placed into the context for every conversation, it would consume a large number of tokens and allow irrelevant rules to interfere with the current task.
Agent Skills use progressive loading:
First layer: name + description
Used for discovery and judgment at startup
Second layer: SKILL.md body
Load the full process after confirming task match
Third layer: references, scripts, assets
Read or run on demand when reaching relevant steps
For example, if a user asks "What's the weather like in Beijing today," the Agent only needs to see the name and description of meeting-minutes, judge that it is irrelevant to the task, and does not need to load the entire set of meeting minutes templates.
When the user provides a meeting transcript and requests minutes generation, the Agent then reads the full SKILL.md.
This approach solves two problems simultaneously:
- Allows the Agent to possess more professional skills
- Avoids carrying all skill content for every task
Complex Skills are not just one Markdown file
When tasks become complex, more resources can be added to the Skill directory:
skill-name/
├── SKILL.md
├── scripts/
├── references/
└── assets/
Each directory is suitable for carrying different content:
| Directory | Purpose | Example |
|---|---|---|
scripts/ |
Execute deterministic, repetitive processing steps | File conversion, data validation, report generation |
references/ |
Store knowledge that needs to be consulted on demand | API documentation, business specifications, format rules |
assets/ |
Store resources needed for final delivery | HTML templates, images, fonts, document templates |
SKILL.md should be responsible for explaining the overall process and clearly specifying which file to read under what circumstances.
If all materials are piled into the main file, the Skill will load all content as soon as it is triggered, rendering progressive loading meaningless.
Skill is not just a "saved Prompt"
Understanding Skill as "a saved Prompt" can help beginners get started, but this statement is incomplete.
The core of a simple Skill might indeed be just a piece of structured instruction; complex Skills can also include scripts, reference materials, templates, and evaluation cases, and can be version-controlled along with the project.
A more accurate understanding is:
Prompt: Instructions for this specific task
Skill: A reusable execution manual and supporting resources for a class of tasks
A Prompt might tell the Agent to "generate a meeting minutes," while the Skill is responsible for answering:
- What content should be deleted
- What information must be retained
- How to identify meeting conclusions
- What fields action items contain
- How to handle missing information
- What structure to use for final delivery
What is the difference between Skill, Memory, MCP, and Tool?
These concepts often appear together in Agent applications, but they solve different problems.
| Capability | Problem Solved | Meeting Minutes Scenario |
|---|---|---|
| Prompt | What the user wants to do this time | "Organize this meeting transcript" |
| Memory | What the Agent should remember long-term | Team name, user preferences, fixed terminology |
| MCP | How to standardize connections to external systems | Connect to cloud drives, document systems, or databases |
| Tool | What operations the Agent can perform | Read recording files, save minutes |
| Skill | How this type of task should be professionally completed | Clean, categorize, extract action items, apply template |
MCP and Skill are not substitutes for each other.
MCP / Tool provides "what can be done"
Skill provides "how it should be done"
For example, MCP can allow the Agent to read an enterprise document library, and the meeting minutes Skill then guides the Agent on how to process the read transcription content. One provides connection and operation capabilities, the other provides a stable professional process.
What tasks are worth encapsulating into a Skill?
Tasks suitable for creating a Skill typically have the following characteristics:
- Will be executed repeatedly
- Processing steps are relatively stable
- Have clear requirements for output format
- Have quality standards that are easy to miss
- Team members need to use the same set of rules
- Results can be verified through cases or assertions
Meeting minutes, AI daily reports, code reviews, weekly report generation, and brand document creation all fit these characteristics.
The following situations may not require creating a Skill:
- One-off temporary questions
- Open-ended discussions without a stable process
- Tasks that can be stably completed with a simple one-line Prompt
- Work where rules are still changing frequently and consensus has not yet formed
The goal of Skill is not to turn all conversations into processes, but to solidify the parts that are truly stable, repetitive, and have professional requirements.
Create your first Skill starting from a repetitive task
A feasible creation sequence is:
Find a recurring task
→ Clarify input and output
→ Write out professional processing steps
→ Create SKILL.md
→ Prepare several sets of real test tasks
→ Check triggering and output effects
→ Continue modifying based on failure results
Finishing writing the file does not mean the Skill is complete.
The description might not cover the expressions of real users, the body process might miss edge cases, and a fixed template might cause certain types of meetings to lose important information. Only through repeated verification with different test cases can you know whether this Skill is truly more stable than a regular Prompt.
How to use skill-creator to create a meeting minutes Skill, and complete evaluation and iteration by comparing results with and without the Skill, will be implemented in the next article.
Summary
Agent Skill organizes the instructions, scripts, and resources needed to complete a class of professional tasks into a directory, and loads them dynamically based on user intent.
Its core structure can be summarized as:
name: Who it is
description: When to use it
SKILL.md body: How to complete the task after triggering
scripts / references / assets: Resources used on demand during execution
Skill does not magically grant the model new knowledge, nor does it replace MCP and Tool. It turns the working methods originally scattered across chat histories, personal experiences, and ad-hoc Prompts into reusable, maintainable, and evaluable Agent capabilities.
References
- Anthropic: Skills Official GitHub Repository
- Agent Skills: Specification
- Anthropic Engineering: Equipping agents for the real world with Agent Skills