Skills

A skill is a package of task instructions for an agent. It can include instructional guidance, reference material, templates, assets, and scripts. It tells an agent how to approach a class of tasks, but installing or selecting it does not automatically grant file, command, network, or MCP permissions.

Three sources

A workspace skill comes from editable project content and is treated as untrusted instruction by default. An installed skill has been confirmed by the user, but “user installed” still does not mean that every script in it is safe.

Select a skill for a task

1. Create or open a conversation.
2. Select “+” in the composer.
3. Select “Skills.”
4. Search for and select the skills you need.
5. Verify the source label, then send the task.

A run can currently select or activate at most eight skills. More is not necessarily better: irrelevant instructions consume context and can introduce conflicts. Usually, select only the one or two skills closest to the goal.

If a selected skill changes before you send, the interface requires “Use latest version.” If the skill is removed or you leave its project, remove the invalid selection before sending.

An agent can also discover and activate a matching skill from the currently available catalog. Activation lasts for the current run only and does not make the skill a permanent part of the conversation.

Install a third-party skill

1. Open “Settings → Skills.”
2. Select “Install skill.”
3. Select “Install from GitHub” or “Install from local folder.”
4. GitHub sources can be a public repository home page, a directory within a repository, or an exact SKILL.md link. If a source contains multiple skills, select one.
5. Wait for Captain Who to inspect the package contents.
6. Review the source, file count, scripts, changes, and security warnings, then confirm installation.

An installation preview expires and must be inspected again afterward. After installation, you can enable, disable, update, or uninstall the skill from the list.

Uninstalling removes only the local skill package managed by Captain Who. It does not modify the original GitHub repository or the source directory you selected.

Minimal workspace-skill structure

A minimal workspace skill has this structure:

.agents/
└── skills/
    └── my-skill/
        └── SKILL.md

The entry file must be named exactly SKILL.md. A skill can also include resources such as references/, assets/, templates/, and scripts/. To create a complete skill, see the Create a Skill Tutorial. Do not directly modify built-in skills installed with the app.

Special boundaries for skill scripts

Skill Scripts currently support only Python scripts contained in the package and depend on a verifiable local Python 3 installation. Running a script requires Full read/write scope and the Full Access prerequisite, and every execution requires explicit approval. Missing Python dependencies are reported rather than installed automatically.

In other words, “the script comes from a skill” is not a reason to run it automatically. At approval time, still inspect the skill, script, arguments, dependencies, and target project.

Built-in Office skills

Creating or editing Word, Excel, and PowerPoint files depends on the corresponding built-in skill. The lower-level Office tools primarily inspect, validate, and render. Actual creation and editing use the managed Builder/Editor supplied by the skill.

See Office, PDF, Spreadsheets, Presentations, and Images for detailed steps.

Current limitations

• Third-party installation sources are primarily public GitHub repositories and authorized local folders. There is no general skill store.
• Activation lasts for one run only; a skill cannot be permanently enabled for a conversation.
• Very large skill directories are trimmed and display diagnostics.
• A skill is not an operating-system sandbox, and other programs may modify the project directory at the same time.

Related topics

• The Difference Between Skills and MCP
• Create and Use a Skill
• Permissions and Approvals