Skills System
Create, import, version, and bind reusable operating guides for agents.
A skill is a reusable operating guide. It can be a single SKILL.md file or a complete bundle with references, scripts, assets, and data files. After you bind a skill to a custom model, the agent sees its name and short description and loads the detailed instructions only when the task needs them.
Skills and system prompts serve different purposes. A system prompt applies to every conversation. A skill is better for a workflow, rule set, or example collection needed only for particular tasks.
With Agent Core 1.0, the names and descriptions of bound skills are included in the agent's working information. The model calls load_skill when it needs the full instructions. Write a description that clearly states when the skill applies instead of using a vague feature label.
Create a skill
In the sidebar, go to My assets → My skills.
Select New skill, enter a name and short description, then select Create skill.
Write the SKILL.md content in the editor. Saving edited content creates a new skill version.
A useful SKILL.md normally includes:
- When to use the skill and when not to use it
- Steps and required checks
- Input and output formats
- Error handling and limitations
- Short, concrete examples
Import a skill bundle
On My skills, select Import SKILL.md to upload a Markdown file or skill bundle. The supported formats are .md, .zip, and .skill.
A bundle can contain:
my-skill/
├── SKILL.md
├── references/
├── scripts/
├── assets/
└── data/Relative paths in the bundle are preserved. Agent Core 1.0 primarily loads the skill text on demand. Agent cores that provide a file-backed skill workspace can also use the supporting files.
Versions and review
Each content update can become a new version. A custom model is bound to the skill itself, and the version it actually uses is resolved on every read, so you never have to re-bind after an update:
| Skill source | The custom model uses |
|---|---|
| A skill you created | Your latest saved content, even before it is submitted or published |
| A skill adopted from the Marketplace | The publisher's latest approved version; their unpublished drafts never reach you |
Your own skills can be tested in Playground the moment you save them — no version bump required. Someone else's skill only affects your model once they publish it.
| Status | Meaning |
|---|---|
| Draft | A version still being edited |
| Pending | Submitted for Marketplace review |
| Approved | Available to eligible users |
| Rejected | Not approved; it can be revised and resubmitted |
Marketplace review and model binding are separate workflows. Owners can bind their own skills to a custom model. When adopting another publisher's Marketplace skill, only approved skills that you are allowed to use are selectable.
An imported or adopted skill may be read-only. Use Copy as my skill to create an editable copy, or fetch a newer approved version if you want to continue following the publisher's releases.
Bind a skill to a custom model
Open the custom model's Customize page.
Go to Behavior → Skills.
Select the skills you want from My skills or from the Marketplace skills available to you.
Save the configuration. New chat requests use the updated bindings.
The selectable skills depend on ownership, organization scope, and Marketplace entitlement. A user without permission to edit the custom model cannot change its bindings.
Runtime notes
- A skill is not loaded on every turn. The model decides from the task and skill description.
- Loaded instructions consume model context. Keep the main workflow in
SKILL.mdand place large supporting material in reference files. - A skill provides instructions; it does not grant database, API, or MCP permissions.
- When Agent Core memory is enabled, a background review may selectively save a reusable method as a skill. Most ordinary conversations create nothing.
- After editing your own skill you do not need to re-bind it. The next new chat request picks up the latest content; requests already in flight keep the version they started with.