Skip to main content

Skill Management

A Skill is a pluggable capability unit of Super Agent. Administrators can centrally manage organization-level Skills on the platform to provide reusable capability modules for Super Agent.

When a user request involves a specific format, fixed workflow, or domain convention, Super Agent will prioritize routing to the corresponding Skill instead of temporarily organizing rules.

Skill Types​

Skills are divided into two categories:

  1. Prompt-based Skill: contains only SKILL.md, without execution scripts; essentially a structured system prompt.
  2. Execution-based Skill: in addition to SKILL.md, also includes execution metadata such as scripts/, manifest.json, and skill.runtime.json.

Management Entry and List View​

  1. Go to the skill management page: navigate to Organization Studio -> Skills Management.
  2. The system displays the name and description of all uploaded skills in card form.

Common skill package examples:

  • pdf: read, parse, or process PDF files.
  • xlsx: read and write spreadsheets and process structured data.
  • slides-json-generator: generate JSON data required for PPT/slides.
  • Chart Generator: generate charts (bar charts, line charts, pie charts, etc.) based on input data.

Add Skill Package​

Skill management supports uploading skill packages from local files to extend organizational capabilities.

  1. On the skill management page, click Create.
  2. Select a local file or drag and drop to upload directly.
  3. After the upload is complete, return to the skill list to confirm the import was successful.

Tip: After a skill is uploaded, it must also be associated in the Super Agent management configuration before it can be automatically recognized and invoked at runtime.

Upload Limits​

  • Number of skill packages per upload: 1
  • Supported skill package formats: .zip / .skill / .md
  • Skill package size limit: 50 MB

Skill Configuration​

After importing a skill, you can adjust the configuration according to business needs.

  1. Select the target skill from the skill list.
  2. Click ✏️ to enter the settings page, where you can re-upload the skill package.

Platform Runtime Capabilities and Boundaries​

Automatically Handled by the Platform​

  • Environment variable injection (SKILL_DIR, SKILL_OUTPUT_DIR, etc.)
  • Bare filename resolution (prefers workspace/, then historical runs/)
  • Known path parameter rewriting (from skill.runtime.json)
  • Working directory switching (skill / workspace / output)
  • Automatic artifact detection based on output_patterns and manifest

What the Platform Cannot Safely Infer​

  • Which positional arguments are file paths (when the Skill does not declare path_args)
  • Whether a third-party CLI returns misleading exit codes
  • Whether the script should read from workspace, the previous run directory, or resources inside the Skill package
  • Which generated files are final deliverables (when output_patterns or manifest outputs are not declared)

Security Boundaries​

  1. Must run in the platform sandbox; do not assume the script runs on the business host machine.
  2. Skill self-service dependency installation is not supported; it does not rely on requirements.txt / package.json.
  3. For additional dependencies, contact the platform for preinstallation.

The script execution methods are as follows:

LanguageExecution MethodRequirement
.pyThe platform loads and executes via runpy.run_path()No special requirements
.sh / .bash / .js and other languagesRun directly as an executable file (subprocess.run([script]))Must have a shebang (such as #!/usr/bin/env bash) and executable permission

Note: The platform does not automatically select an interpreter based on the file extension. If a non-Python script lacks a shebang or executable permission, it will fail to run directly. It is recommended to use a Python script as the entry point and let it invoke other languages.

Relationship with Other Platform Capabilities​

  • Based on the current Skill positioning and security model, direct access to the platform database and platform APIs through Skill is not supported.
  • When accessing external service interfaces, MCP access is recommended first.
  • Skill focuses more on orchestration of “whether to invoke” and “when to invoke,” and does not replace MCP’s authentication management and access constraints.
  • If it is necessary to guide Super Agent to send curl requests through Skill, an access allowlist must be configured in the sandbox.

Sandbox Preinstalled Dependencies​

The Universal Sandbox image has some dependencies preinstalled. Skill scripts can directly use them via import / require without additional installation.

Image source files: serviceme-sandbox/docker/images/universal-sandbox/