Skills

Skills give agents task-specific instructions. VC includes its own design skills and lists the skills each installed provider reports. Where the provider supports it, you can turn its skills on or off.

Open Settings → Skills and pick a device. Skills are grouped by provider. Use the search field to filter by name, description, or scope; a skill’s scope tells you whether it comes from your user config, the current project, a repo, a plugin, or system policy.

VC design skills

VC includes interface design skills that are added and improved through VC updates. Ask an agent in VC Chat or a CLI agent opened from VC to list the available skills or use one, for example: “Use VC’s Light Ops Sidebar skill for this page”, or ask it to choose a skill for the interface you are building.

The available catalog follows the version of the connected VC server. Updating the desktop app updates skills for its bundled local server. For a remote environment, update VC on the host machine to get new skills. After the updated server starts, open a new Chat or CLI agent session to use the current catalog.

These skills are loaded on demand through the session’s VC connection. They work with web, desktop and mobile clients, including remote environments, and remain available when browser access is turned off. Closing a CLI agent revokes its connection without affecting the chat session. They are not installed in your provider’s global or project skill directories, and do not appear in the provider skill list in Settings or the $ menu. Agents launched outside VC, including clients using an external MCP token, cannot load them.

CLI agents with native MCP support can call skills_list and skills_read directly. Other CLI agents can read VC_MCP_INSTRUCTIONS in their environment for the same commands through VC’s authenticated bridge.

Managed Computer Use skill

The Computer Use section at the top of the page manages a skill VC owns: it installs vc-computer-use into the device’s cross-agent .agents/skills directory plus any provider skill directories that already exist (~/.claude, ~/.codex, ~/.gemini). The skill teaches agents the inspect-before-act workflow for the computer_use MCP tool. The tool itself still requires the Computer Use switch in Settings → Integrations.

VC only writes copies it owns. A hand-written vc-computer-use is left in place when you turn the skill on, and turning it off removes only copies that carry VC’s managed marker. Installing an agent later adds a skill directory VC did not write to, so the row can report fewer directories than it lists; the server fills the gap and refreshes the skill text on its next start.

Bundled skills

On start, the server also installs a set of skills into ~/.agents/skills and the Claude, Codex, Cursor, and Gemini skill directories: database, docker, the vc-* guides for VC’s own tools, and the Ponytail skills. Ponytail steers agents toward the smallest change that works: reuse before rewrite, standard library and platform features before new dependencies, without cutting validation, error handling, security, or accessibility.

• ponytail applies to coding tasks. Say /ponytail lite or /ponytail ultra to change how hard it pushes, or “stop ponytail” to turn it off for the conversation.
• ponytail-review reviews a diff for over-engineering only, and ponytail-audit does the same for the whole repository.
• ponytail-debt lists the ponytail: comments agents leave on deliberate shortcuts.

Deleting a bundled skill does not stick, because the server writes it again on its next start. Turn it off below instead, or put your own SKILL.md with the same name in its place; VC never overwrites a file it did not write.

Enable and disable

• Claude: toggling writes skillOverrides in that instance’s user settings file (settings.json in its config directory, usually ~/.claude). Turning a skill off writes "off"; turning it back on removes the entry, so the skill’s own frontmatter decides again. Claude re-reads the file on each message, so the change applies to your next send.
• Codex: toggling asks the Codex app-server to write its own skill config. A session that is already running may keep the skills it started with until a new session is opened.

After a change, VC refreshes the provider snapshot and any cached project skill list, so the composer $ menu reflects it.

Skills you cannot toggle show a disabled switch:

• VC does not write that provider’s skill config yet (Cursor, Grok, OpenCode, and Antigravity today). Use the provider’s own CLI or configuration for those.
• A higher-priority file owns the skill — for Claude, a project, repository, or administrator-managed settings file outranks your user settings, so writing there would not change anything.

Project-scoped skills are only listed for the device that owns the project. To manage a project’s own skill files, edit the files the provider reads (.claude/skills, .agents/skills, and equivalents).