Skills
Give agents reusable expertise: when to use a skill, how agents decide to use one, running a skill by name, where skills come from, and creating your own.
A skill gives agents reusable expertise: a style guide, a review checklist, the steps to turn a dataset into a figure. It’s a folder with a SKILL.md file, and agents load it only when a task calls for it. Create one when you keep pasting the same instructions into chat, or when a section of your AGENTS.md has grown into a step-by-step procedure.
This page covers how skills work with agents. For installing skills from the marketplace and the full authoring guide, see the Skills section.
When to use a skill
| You want | Use |
|---|---|
| Rules that apply to every chat in one project | AGENTS.md |
| Expertise or a procedure the AI should use only when a task needs it, in any project | A skill |
| A worker you run in one click, with its own job and MODES | A custom agent |
| A single fact the AI should remember | Agent memory |
| A new editor, viewer, or command in the app | An extension |
These combine: an agent can call on skills, so a custom “Methods reviewer” agent can use your “methods checklist” skill.
How agents use skills
- Only the description is read at first. The AI reads each enabled skill’s description and decides, task by task, whether it’s relevant.
- The instructions load when needed. Only then does the AI load the skill’s full instructions, so skills don’t crowd unrelated work.
- Declared tools are respected. A skill that lists the tools it may use is held to exactly those tools while it runs.
- You ask for the outcome. You don’t have to pick a skill — describe what you want and the AI brings in the right expertise. You can also run one by name.
Run a skill by name
Type / in the chat box to open one menu for commands, skills, and saved prompts. It filters as you type. Run a skill by its name:
/literature-reviewRunning a skill by name is useful when you want to be sure it’s used — for example, a house style you always apply before submission.
Where skills come from
| Source | How you get it |
|---|---|
| The marketplace | Browse the Skills catalog, choose Install to open it in Multilo, then enable it in the Skills Manager. See using skills. |
| Your own | Create one with /createskill, from the Skills Manager, or by hand. |
| Other AI tools | Skills in .claude/skills or .agents/skills folders also load, so skills you wrote for other tools can work here too. |
| A colleague | Import a .skill.json bundle they exported. |
Create a skill for your agents
Start the skill
Type/createskilland describe what the skill should do — Multilo creates the folder, the frontmatter, and a first draft of the instructions. Or create the folder andSKILL.mdyourself.Write a description that says when to use it
The AI picks a skill from its description alone. Lead with the task and the cues that should trigger it — “Use when…”.Keep the instructions short and concrete
Write the steps or rules the agent should follow, and say what it should not do.Lint and test it
Fix anything the built-in linter flags, then ask for a task that should use the skill and check the result.
For example, a skill that checks a Methods section:
---
name: methods-checklist
description: >-
Check a Methods section against a reporting checklist. Use when the user
asks to review, check, or tighten a Methods or Methodology section.
allowed-tools: ["editor.read"]
---
# Methods checklist
When reviewing a Methods section, check that it states:
1. The design, and why it fits the research question.
2. Participants or data: who, how many, and how they were selected.
3. Measures and instruments, with their sources.
4. The procedure, in the order it happened.
5. The analysis, including software and tests used.
List anything missing, with the paragraph it belongs in.
Don't rewrite the section unless asked.A skill can also bundle helper scripts and reference files in its folder. See authoring skills.
Frontmatter fields
| Field | What it does |
|---|---|
name | Identifies the skill, and is what you type after / |
description | What the skill does and when to use it — the AI decides from this |
allowed-tools | The tools the skill may use while it runs |
Two kinds of skill
- Reference skills hold knowledge the AI applies while it works — a journal’s house style, your field’s terminology, a citation style’s fine print.
- Task skills hold a procedure — check a Methods section, build a literature matrix, turn a results file into a figure. These are the ones you’ll often run by name.
Either way, keep the instructions concise: when a skill is used, its instructions become part of what the model reads, and that uses credits.
Manage your skills
The Skills Manager lists every skill you have, installed or written yourself. Enable, disable, edit, or remove each one there. Skills travel with the app, so they’re available in every project.
Trust and safety
- A skill can’t quietly reach for tools it didn’t declare.
- Scripts a skill runs use the same secure, permissioned environment as other agent code.
- Marketplace skills are reviewed and signed, and show a trust tier — first-party, verified publisher, or community — before you enable them.
Read before you enable
A skill shapes what the AI does. Read the instructions of a skill from someone you don’t know before you turn it on. See security.Share a skill
- With a colleague: export the skill as a
.skill.jsonbundle and send it; they import it on their side. - With everyone: publish it to the marketplace through the same reviewed, signed path as extensions. See publishing skills.
Troubleshoot
| Problem | What to do |
|---|---|
| The AI never uses my skill | Check it’s enabled in the Skills Manager, and make the description say exactly when to use it. Or run it by name with /. |
| The AI uses my skill when it shouldn’t | Narrow the description, so it names the specific task and cues. |
| The skill can’t do part of its job | Add the missing tool to allowed-tools, then lint it again. |
The skill doesn’t appear in the / menu | Make sure it’s installed and enabled, and fix any problems the linter reports. |