Skills

A skill is a reusable workflow written in Markdown. Use one for recurring work such as reviewing a change, diagnosing a desktop issue, or checking a release. It gives the agent a task-specific procedure; it does not guarantee a correct result or grant extra permissions.

Find and use a skill

In the desktop composer, enter /skills to open the catalog. Search the discovered skills, inspect their source, and use Try now for a user-invocable skill. Skills also appear in the slash menu, where you can pass arguments after the name:

/release-check 2026.9.1

Selecting a skill from the menu or catalog inserts a /skill draft with an explicit identity. Keep that identity intact and add your request below it. Project-local selections bind the originating workspace and a repository-relative instruction file; the selected conversation checkout supplies that file’s version and resources, including the first turn of a new worktree. Browsing the catalog does not create a worktree. User, plugin, and other external selections retain absolute-path identities. On submission, Wuu matches the identity against the discovered, permitted catalog before calling the model, including skills hidden from automatic selection. Missing, out-of-scope, or cross-project selections are rejected; existing absolute-path drafts are never silently remapped. Bare built-in commands such as /review and /compact retain their usual behavior.

The Wuu engine gives the model a catalog of skill names and descriptions. When a task matches, it can use load_skill to read the full workflow. A missing description keeps a skill out of that automatic-selection catalog. disable-model-invocation: true also prevents the model from loading it by name, while user-invocable independently controls direct user invocation.

Availability depends on the workspace, discovered sources, and current tool surface. A skill that requires unavailable tools may be filtered out. Requirements can name reachable built-in, plugin, or MCP tools by their exact published names, including deferred and PTC tools. Session tool restrictions apply to those requirements even before a tool is discovered or loaded. They never enable a disabled tool or grant permissions. Inspect the source path if the wrong version appears.

Twin skills

Twin’s main and reply composers expose user-invocable skills through /. Type to filter, use Up/Down and Enter or Tab to select, then add your request and send. Escape closes the menu; Shift+Enter adds a line. /settings opens Twin’s existing settings panel without sending a message.

Install a Twin skill at $WUU_HOME/user-spaces/<space>/personal/files/.agents/skills/<name>/SKILL.md. The default local Twin also discovers Coding’s user-level skill directories. A Twin skill takes precedence over a shared skill with the same name. Project skills remain scoped to their projects, and shared skills run with Twin’s own tools and permissions. Other user spaces and verified isolated hosts only discover their own Twin skills. The existing skills.enabled setting must be enabled. Reopen Twin to refresh the menu after installing a skill; submission always rechecks its identity and current availability. Discovery failures retain the draft and offer Retry.

Selected skills are expanded only for authenticated user messages. Their contents are saved when the message is accepted, so queued messages and receipt retries retain the same instructions and attachments. Plugin wakeups remain literal. Unknown commands, file paths and multiline text remain ordinary messages.

Editing a local skill

Project and user skill edits do not require a runtime restart. Browsing rereads the local catalog with a short cache; model rounds adopt changes at the next safe step boundary, adding a catalog update only when its visible entries change. Loading a local skill reads the current instructions from that admitted file and rechecks invocation and tool requirements. A missing, renamed, retargeted, or malformed file cannot silently supply stale instructions or switch that load to another same-name source.

Instructions already loaded into history, including explicit selections submitted to a queue, keep their original contents. Plugin skills remain part of their pinned package generation and use the extension update lifecycle. Local discovery preserves the documented source precedence when a later step adopts a new catalog.

Trust the workflow before using it

Read a third-party skill and its supporting scripts before enabling its workflow. The normal load_skill path substitutes arguments and loads instructions; it does not run inline shell expressions or code fences at load time. The instructions can still ask the agent to run commands or access the network afterward, subject to the session’s tool and permission boundaries.

Project instructions such as AGENTS.md apply continuously within their scope. A skill is loaded for a particular workflow, while memory stores information worth keeping across tasks.

Create or install one

Place a skill in a project to share it with contributors, or in a user skills directory to reuse it across projects. The authoring guide covers discovery order, supported metadata, arguments, and validation.

wuu skills lint path/to/skill
wuu skills lint --json path/to/skills-root

Lint checks whether the structure and metadata can be loaded. Review the workflow’s behavior with a low-risk task as well.