Skills

Instructions loaded on need¶

A skill is one instruction file that gives Mewbo a specific way of working. Each lives in its own directory as a SKILL.md file, with YAML frontmatter and a markdown body. Mewbo pulls that body into context only when the skill activates, so dozens of installed skills cost nothing until one runs.

Writing a skill¶

Create a directory at .claude/skills/<your-skill-name>/SKILL.md. The file starts with a YAML frontmatter block and the instruction body follows it.

---
name: code-reviewer
description: Review code changes for correctness, style, and test coverage. Use when asked to review a diff, PR, or commit.
allowed-tools: read_file aider_list_dir_tool aider_shell_tool
---

# Code Review

1. Read the diff using `read_file`.
2. Run `aider_shell_tool` with `git diff HEAD~1` to verify context.
3. Check for missing tests.
4. Report each finding as **Issue**, **Severity**, **Suggestion**.

Frontmatter reference¶

Tool scoping¶

Set allowed-tools to narrow the tool set for as long as the skill is active, using the same tool IDs as anywhere else in Mewbo.

allowed-tools: read_file aider_list_dir_tool

Omit it and the skill inherits the session's full tool set.

Capability gating¶

The client advertises what it can render on the X-Mewbo-Capabilities header, and SkillRegistry filters against it. A gated skill reaches neither the model nor an explicit /skill-name call.

The built-in widget-builder plugin works this way. Its st-widget-builder skill declares requires-capabilities: [stlite], so it never appears on a CLI session that does not advertise stlite.

---
name: st-widget-builder
description: Build an interactive stlite widget rendered inline in the console.
requires-capabilities: [stlite]
---

Declare requires-capabilities once in plugin.json and every skill under that plugin inherits it. See Plugins & Marketplace → Capability gating.

Shell preprocessing¶

Skills embed shell commands with the !`command` syntax. Each matched command runs at activation time and its standard output is substituted inline before the model sees the instructions.

Example. Inject the current git branch name.

You are reviewing code on branch: !`git rev-parse --abbrev-ref HEAD`

A command times out after 30 seconds. On error the placeholder becomes [ERROR: ...].

Where skills live¶

Plugins ship skills in the same SKILL.md format. A plugin skill never overrides a personal or a project skill of the same name.

Invoking skills¶

Two paths trigger a skill.

• Automatically. Mewbo reads the skill catalogue at the start of every session and can activate a matching skill from your request.
• Explicitly. Type /skill-name in the CLI or console. Arguments after the name arrive as $ARGUMENTS inside the skill body, and individual tokens as $0, $1 and so on.

Sub-agents carry the same catalogue in their system prompt, so naming a skill in a delegated task description works too.

Hot-reload¶

A changed SKILL.md is picked up automatically, and a new skill directory on the next scan. No restart is required.