Skills

Skills give Qwen Code reusable instructions and supporting files for a particular task. A skill can be selected with a slash command or made available for the model to select from its description. Installing a skill makes it discoverable; it does not run its scripts immediately.

Create a project skill

Create .qwen/skills/review-tests/SKILL.md:


---
name: review-tests
description: Review a change for missing tests and explain the most important gaps.
argument-hint: "[file or directory]"
disable-model-invocation: true
---
Review the tests relevant to $ARGUMENTS.
Read the implementation and existing tests before making recommendations.
Report concrete missing cases with file paths. Do not edit files.

Start Qwen Code in the project and enter /review-tests src/. Use /skills to inspect available skills and their enabled state. The directory name determines this skill’s command identity. Keep name aligned with it for clarity.

Place supporting material beside SKILL.md, for example references/checklist.md or scripts/check.py, and describe when to read or run it in the skill body. A script needs its own interpreter and dependencies. A skill is instructions, not a sandbox.

Locations and visibility

Location	Use
`~/.qwen/skills/<name>/SKILL.md`	Personal skills across projects
`.qwen/skills/<name>/SKILL.md`	Skills shared with a trusted project
Plugin `skills/` directory	Skills supplied by an enabled plugin

The user root follows --config-dir when configured. Additional workspace directories may also contribute project skills. Project trust, disabled skills, administration settings, and name conflicts can affect which skills are available. The optional .agents/skills compatibility directory is controlled by configuration; it is not required for Qwen Code skills.

Choose unique command names. Category folders under skills/ are scanned, but a skill is named after its own folder: .qwen/skills/team/review-tests/SKILL.md is /review-tests. Colon-separated names apply to files under .qwen/commands/ and to plugin skills, which are prefixed by the plugin name. Use the displayed name in /skills, particularly when more than one source defines a skill.

Frontmatter

Field	Meaning
`description`	Explains when the skill is useful; write a specific task trigger
`argument-hint`	Hint displayed while entering arguments
`arguments`	Array of `{name, description, required}` argument descriptions
`disable-model-invocation: true`	Prevents autonomous model selection; useful for explicit workflows
`user-invocable: false`	Hides the skill from user slash invocation; model selection is controlled separately
`allowed-tools`	Tool names or rules used by the skill’s execution policy
`model`	Model override; the model must be available in the session’s configuration
`effort`	`none`, `low`, `medium`, `high`, `xhigh`, `max`, or a positive integer; effective support depends on the model
`context: fork`	Runs the skill in a separate agent context
`agent`	Agent used with a forked skill
`hooks`	Hooks associated with skill execution
`paths`	File patterns used for contextual skill visibility
`shell`	`bash` or `powershell` for shell preprocessing

Omit fields you do not need. Tool allowances do not remove project trust requirements or higher-level restrictions. Model selection and user invocation are independent: setting both visibility controls can make a skill unavailable through either normal entry point.

The body supports $ARGUMENTS for the complete argument string, $ARGUMENTS[0] or $0 for the first argument, and subsequent zero-based indices. Quoted arguments remain grouped. Missing indexed arguments become empty strings. When a nonempty argument string has no placeholder, it is appended to the body. ${QWEN_SKILL_DIR} identifies the skill directory for supporting files.

Install and manage


qwen skills list
qwen skills list --all
qwen skills install ./review-tests --scope project
qwen skills link ./review-tests --scope user
qwen skills disable review-tests --scope project
qwen skills enable review-tests
qwen skills uninstall review-tests --scope project

Install accepts a Git repository URL, an HTTPS ZIP URL, a local directory, or a local .skill/.zip archive. Use --path for a subdirectory in a repository or archive. Installation and linking default to user; project and workspace name the same skill scope. Disable defaults to the workspace. Enable does not take a scope flag and clears the applicable disabled entries.

Linking is useful while developing a local skill because it references the source. Review instructions and scripts before accepting installation. --consent skips the installation confirmation and should only be used after that review.

Troubleshooting

Missing from /skills: check SKILL.md, YAML syntax, project trust, the scope, and loading diagnostics. Keep SKILL.md inside its own named directory.
Listed but absent from slash completion: check user-invocable, disabled state, and the full namespace.
The model never selects it: check disable-model-invocation, its description, and any paths filter. Test explicit invocation first.
An edit is not reflected: skill files are watched, but deep asset changes may not trigger a refresh. Restart the session after changing nested support files.
A tool or model is unavailable: verify it independently in the session; frontmatter cannot install missing tools or grant an account access to a model.

See custom commands, plugins, and hooks.