Create a skill
A Skill defines reusable capabilities for an Agent, such as script execution, data processing, and external API calls. Create a Skill online with Markdown or by uploading a skill package. Agents can then call the Skill at runtime as needed.
Overview of creation methods
|
Creation method |
Description |
Use case |
|
Define a Skill online using Markdown. |
Quickly create lightweight Skills, with support for AI generation. |
|
|
Upload a local file or import a skill package from a URL. |
Import and use existing, fully developed skill packages. |
Method 1: Define with markdown
Use Markdown to define a Skill's description, features, and usage instructions. The platform provides an Insert Example template and an AI Generation feature to simplify authoring.
Procedure
-
Log in to the AgentRun console.
NoteBefore logging in, ensure you have activated Function Compute and have the required permissions for AgentRun. If not, go to the Function Compute product page to activate the service.
-
In the left-side navigation pane, choose Tools and Skills.
-
Click Create Tool.
NoteThe button on the list page is labeled Create Tool, but the title of the creation page is "Create Tool or Skill".
-
In the Creation Type section, select Skills.
NoteThe Creation Type section provides the following options:
-
Skills: Create or import a skill package. Supports multiple formats such as Markdown and code packages.
-
MCP: An MCP (Model Context Protocol) tool. Supports remote connections and local deployments.
-
Function Call: Create a Function Call tool. Supports formats such as OpenAPI and RESTful APIs.
-
Tool Marketplace: Select an existing tool from the marketplace.
-
-
In the Creation Method section, select Define with Markdown.
-
Configure the following parameters.
Parameter
Required
Description
Skill name
Yes
The unique identifier for the Skill.
Description
No
A brief description of the Skill's purpose. The maximum length is 256 characters.
Markdown content
Yes
Use Markdown syntax to define the Skill's behavior. For detailed guidance, see Write Markdown content below.
Credentials
No
Configure how the Skill authenticates at runtime. The following modes are supported:
-
Use existing credentials (Default): Reuses the credentials already configured for the current account to ensure data security.
-
Anonymous access: Does not send authentication information. This mode is suitable for accessing public resources.
WarningAnonymous access does not guarantee data security. Before selecting this option, confirm that the target service supports unauthenticated calls.
For guidance on writing the Markdown content, see Write Markdown content below.
Use AI generation
If you are unsure how to write the Markdown content, you can use the AI Generation feature:
-
In the upper-right corner of the Markdown content section, click AI Generation.
-
Enter a natural language description of the Skill's functionality.
-
The AI automatically generates a Markdown definition that conforms to the required format.
-
Once generated, you can modify and refine the content in the editor.
NoteThe AI-generated content serves as a starting template. Adjust it to your specific needs before submission.
Content security scan
When you finish writing, click Scan Markdown Content below the editor to check for malicious code or scripts.
-
-
(Optional) Configure the workspace.
In the workspace section, select the workspace for the Skill.
-
Click Create Now.
Write markdown content
In the Markdown content editor, use Markdown syntax to describe the Skill's capabilities. Include the following information:
-
Title: The Skill name.
-
Description: A summary of the Skill's functionality.
-
Skill list: The specific functions included in the Skill and their purpose.
-
Usage: How to call the Skill, with code examples.
Click Insert Example in the editor's toolbar to populate a sample template and see the required format and structure.
The following example, generated by Insert Example, shows the recommended structure for a Skill definition. This is not executable code.
---
name: weekly-report-generator
description: An intelligent assistant for generating structured weekly reports. When a user mentions keywords such as "weekly report", "this week's summary", "work progress", or "team update", this Skill should be proactively called, even without an explicit request, to help them quickly produce a well-formatted and focused report.
---
# Weekly Report Generator
This Skill helps users efficiently write professional and clear weekly reports by automatically extracting key information and organizing it into a standard structure. This ensures a logical, complete, and focused report.
## When to use this Skill
- When a user mentions keywords like "weekly report", "this week's summary", "work progress", "team update", "progress update", or "weekly update".
- When a user needs to sync work results and plans with their manager or across departments.
- When a user wants to save time writing without sacrificing report quality.
Avoid using in the following situations:
- The user only requests a simple diary-style record or informal communication.
- Internal chats that do not require formal reporting.
## Core workflow
1. **Identify Report Type**: Determines if it is a personal, team, or project weekly report, as each type has a different focus.
2. **Extract Key Information**: Gathers the following elements by asking questions or analyzing context:
- This week's accomplishments (completed tasks, achieved goals)
- Ongoing work (current progress, blocking issues)
- Next week's plan (to-do items, priority ranking)
- Resources or coordination needed.
3. **Apply Standard Template**: Automatically generates content based on a preset structure, ensuring concise language and clear organization.
4. **Optimize Phrasing**: Uses positive language to describe achievements and explains the reasons for and countermeasures against difficulties to improve readability and professionalism.
## Output specification
### This Week's Work Summary
- [Accomplishment 1] (Briefly describe the result)
- [Accomplishment 2] (Quantifiable results are better, e.g., "completed 3 module developments")
### Currently In Progress
- [Task Name]: Current progress is 70%, expected delivery next Wednesday.
- [Task Name]: Encountered [specific issue], tried [solution], need [resource/assistance].
### Next Week's Work Plan
- [High Priority] Task 1: Goal + Deadline
- [Medium Priority] Task 2: Goal + Deadline
- [Low Priority] Task 3: Planned to start
### Other Notes
- [Collaboration needs, risk warnings, suggestions, feedback, etc.]
## Notes
- Keep each item to 1-2 lines to avoid lengthy descriptions.
- Emphasize the value of achievements, not just the process. For example, "Optimized API response time to 200ms" is better than "Modified the code logic."
- If the user does not provide specific details, ask follow-up questions like, "What are the key achievements you want to highlight this week?" or "Are there any matters that require cross-departmental coordination?"
Method 2: Upload a skill package
Import an existing skill package by uploading the file or providing its URL.
Procedure
-
Log in to the AgentRun console.
NoteBefore logging in, ensure you have activated Function Compute and have the required permissions for AgentRun.
-
In the left-side navigation pane, choose Tools and Skills.
-
Click Create Tool.
-
In the Creation Type section, select Skills.
-
In the Creation Method section, select Upload Skill Package.
-
Configure the following parameters.
Parameter
Required
Description
Skill name
Yes
The unique identifier for the Skill.
Description
No
A brief description of the Skill's purpose. The maximum length is 256 characters.
Credentials
No
Configure how the Skill authenticates at runtime. The following modes are supported:
-
Use existing credentials (Default): Reuses the credentials already configured for the current account to ensure data security.
-
Anonymous access: Does not send authentication information. This mode is suitable for accessing public resources.
WarningAnonymous access does not guarantee data security. Before selecting this option, confirm that the target service supports unauthenticated calls.
-
-
In the Upload Skill Package section, select an import method.
-
Upload File
-
Select Upload File.
-
Drag the skill package file to the upload area, or click the upload area to select a file.
Supported file formats and limits:
Item
Description
File format
ZIP
File size
Up to 50 MB
ImportantEnsure that a valid
SKILL.mdfile exists in the root directory of the ZIP archive.(Optional) Select Enable Security Scan to scan the uploaded skill package.
-
-
From URL
-
Select From URL.
-
Enter the URL of the skill package.
-
The platform will download and import the skill package from the specified URL.
-
-
-
(Optional) Configure the workspace.
-
Click Create Now.
Skill loading and consumption
After a Skill is created, an Agent must load it before use. AgentRun supports the following loading methods.
Proactive loading
When creating an Agent, add a Skill on the Skills tab under Tools & Context. The Skill is then proactively loaded into the Agent's execution environment.
The loaded content includes the Skill's script files, description, and npm dependencies. The system automatically installs all dependencies before execution. Skills are distributed as ZIP packages. The first load may have a short delay due to cold start and package size. Subsequent calls reuse the instance and are significantly faster.
You can select the Allow Skill to Execute Scripts option when creating an Agent to authorize the Skill to execute code at the instance's operating system level. If enabled, scripts included in the Skill will run directly. Only add Skills from trusted sources.
Proactive discovery
AgentRun includes a built-in find_agent_on_skills capability that lets an Agent search for and discover available Skills under the current account at runtime. This capability requires no manual addition.
This capability is triggered only when a user issues an explicit query, such as "List my Skills", and does not search automatically. It also supports dynamic installation of Skills based on task requirements.
Permission control
Role permissions configured in the Agent Runtime determine a Skill's visibility through a resource-level authorization system, with no additional restrictions needed at the conversational layer.
-
If a Role is authorized to access only the resources of a specific workspace,
find_agent_on_skillsreturns only the Skills within that workspace. -
For scenarios that require strong isolation, use the Proactive loading model (pre-selecting and loading Skills) instead of relying on runtime discovery.
Execute a skill in a sandbox
Skills can be loaded and executed in a sandbox environment via the MCP protocol. The Sandbox feature is available in the production environment and currently in canary release.
To integrate, first deploy the MCP Server, then start the sandbox proxy function. The Agent connects to the MCP Server in the sandbox via the MCP Client, which then calls the Skill Loader to load and run the Skill.
The call chain is: Agent → MCP Client → MCP Server (Sandbox) → Skill Loader → CLI Engine → External API
For more information, see Dynamically mount custom Skills for a sandbox.
FAQ
Markdown definition vs. package upload
|
Comparison |
Define with markdown |
Skill package upload |
|
Use case |
Quickly create lightweight Skills. |
Import existing, mature skill packages, or Skills that consist of multiple files or folders. |
|
Editing method |
Online editing with AI assistance. |
Import by uploading a file or from a URL. |
|
Flexibility |
Can be modified online after creation. |
Can be opened, edited, and saved in a Web IDE. |
Using AI-generated definitions
AI-generated Markdown content is a good starting point. Adjust and refine it to match your business needs and ensure the Skill behaves as expected.
Upload permissions
Skill uploads are controlled by RAM (Resource Access Management) permissions. Any user with permission to call the CreateTool API (authenticated with an AccessKey) can upload a Skill, as Skills are a type of Tool created with this API.
No additional review process is required. An uploaded Skill takes effect immediately and is available to other users with appropriate permissions. Security relies on the front-end scanning feature, so select Enable Security Scan when uploading.
Unsupported file format error
Symptom: When you upload a skill package, the system shows "Unsupported file format" or the upload fails.
Possible cause: The uploaded file is not in ZIP format, or its file extension does not match the actual format.
Solution: Confirm the skill package is a ZIP file. If the format is correct but the upload still fails, the file may be corrupted. Try re-compressing and uploading it again.
Exceeding the size limit
Symptom: When you upload a skill package, the system reports that the file exceeds the size limit (50 MB).
Possible cause: The skill package contains unnecessary dependency files, log files, or large resources.
Solution: Check the content of the skill package, remove unnecessary files such as node_modules, log files, or test data, and then repackage it with only the core scripts and configuration files. If the package still exceeds the limit after simplification, consider splitting the Skill into multiple smaller, independent skill packages.
First-time loading delay
Symptom: When an Agent calls a Skill for the first time, the response is slow, and the wait time is significantly longer than for subsequent calls.
Possible cause: A Skill's first load triggers a cold start, which involves downloading the code package and installing npm dependencies. The time required depends on the size of the code package and the number of dependencies.
Solution: This initial loading delay is expected and typically lasts from several to over ten seconds. Subsequent calls reuse the existing instance, resulting in significantly faster response times. If you experience persistent timeouts, check if the Skill's code package size or number of dependencies is too large.