Skills

A skill is a reusable way of doing something: instructions, optional scripts and optional reference files. It does not sit in the context until needed, so you can keep many without slowing conversations down.

What is inside

Part	Purpose
Name / description	Lets the agent decide when to use it
Instructions	Steps, constraints, output format
Scripts	Executables called at runtime when needed
References	Docs, templates, examples read on demand

Create a skill

Open Skills → New, fill in name, description and instructions, then add scripts and references as needed.

Import skills

• From a directory: drag it onto the upload area. A typical skill has SKILL.md plus scripts/ and references/.
• From GitHub: enter owner/repo or a URL, optionally a package name and branch. Good for sharing across a team.

After importing, check:

1. The description states the trigger clearly.
2. Scripts are executable.
3. Referenced files are all present.

Use in an agent

In the Agents editor, tick skills under Skills. Search by name when the list is long.

Back up to GitHub

Skills → Backup syncs the skills directory to a repository:

• Repository (owner/repo) and a token with repo scope
• Branch
• Force overwrite replaces same-named remote files with local ones — confirm before running

Writing a good skill

1. Description = trigger, not implementation. e.g. “Use when the user asks for an Excel file”.
2. Instructions = steps, each verifiable; write “confirm first” where it is ambiguous.
3. Scripts should be idempotent; running twice must not break anything.
4. Put large material in references, not instructions.

FAQ

The agent never uses the skill — check the description contains words users actually say, and that the agent has it ticked.

A script fails — verify the interpreter and dependencies; the sandbox image may not include that runtime.

Files are missing after import — repackage the directory and keep references relative.