Agent Skills

Agent Skills

VT Code supports repository, user, admin, and bundled system skills using the open Agent Skills SKILL.md format.

Discovery

Skills are discovered in this order:

1. Nearest ancestor .agents/skills from the working directory up to the git repository root
2. ~/.agents/skills
3. /etc/codex/skills
4. Bundled system skills

If multiple skills share a name, nearest repository skill wins, then user, then admin, then system. Disabled entries from ~/.codex/config.toml are honored.

Skill Structure

my-skill/
├── SKILL.md        # required
├── scripts/        # optional
├── references/     # optional
└── assets/         # optional

SKILL.md Frontmatter

---
name: my-skill
description: Explain what this skill does and when to use it.
license: Apache-2.0
compatibility: Requires git and network access
allowed-tools: Read Write Bash
metadata:
  owner: platform-team
---

Required: name, description. Optional: license, compatibility, metadata, allowed-tools, disable-model-invocation.

disable-model-invocation: true hides a skill from the model-facing startup catalog while keeping it available for explicit harness activation. Legacy VT Code frontmatter (version, author, when-to-use, model, mode, permissions, etc.) is rejected.

Prompting Behavior

• Explicit mention wins: Use the my-skill skill
• Implicit matching uses description
• Full SKILL.md bodies load only when a skill is selected
• Referenced resources load on demand

Commands

vtcode skills list                 # list skills
vtcode skills info my-skill        # inspect one skill
vtcode skills create my-skill      # scaffold a new skill
vtcode skills validate ./path      # validate a skill
vtcode skills config               # show configured paths

Slash Command Skills

The interactive slash-command surface is also exposed as skills using the cmd-<slash-name> form (e.g. cmd-status, cmd-review). The /status or /review slash command remains the primary interactive alias.

Related

• Getting-Started
• Configuration