Author, version, and distribute MCP agent skills across your team — share reusable tool workflows via config file with full version history.

Starting from a run

The fastest way to create a skill is to start from a run that worked well:

Open a completed run in the Runs panel
Click Save as skill in the run header
Name the skill and add a description
Review the auto-generated prompt template and adjust as needed
Click Save

Prompt templates

Skill prompts support {{variable}} interpolation. Variables become the skill's typed inputs:

Summarize the following code review for {{repo}} PR #{{pr_number}}.

Focus on:
- What changed and why
- Risk assessment
- Suggested follow-ups

Use a professional but direct tone.

With this prompt, Invoked generates an input form with repo (string) and pr_number (number) fields.

Input schema

You can add validation and metadata to inputs:

{
  "repo": {
    "type": "string",
    "description": "GitHub repo in owner/name format",
    "placeholder": "acme/api"
  },
  "pr_number": {
    "type": "number",
    "description": "Pull request number",
    "minimum": 1
  }
}

Tool selection

Choose only the tools your skill needs. Fewer tools = faster, cheaper, more focused runs. A summarization skill probably only needs fetch_url. A code-generation skill might need read_file, write_file, and bash.

Testing a skill

Use the Test button in the skill editor to run it with sample inputs. The test run appears in your Runs panel with a [test] tag.

Team distribution

Export a skill as a JSON file and share it:

Open the skill in the Skills panel
Click ⋯ → Export
Share the .json file with your team

Your teammates import it via Skills → Import.

Future versions will support workspace-level skill libraries that sync automatically.

Versioning best practices

Bump the version when you change the prompt significantly
Use the description field to explain what changed
Keep old skills around — don't delete them, archive them
Use evals to verify that a new version is an improvement

Building Skills