On this page

A reusable way of doing the work

An Agent Skill is a directory containing a required SKILL.md file and optional supporting resources. Its metadata helps an agent discover when the procedure applies; its body describes how to complete that kind of work.

Skills sit between an agent and its tools:

LayerResponsibility
AgentIdentity, model, role, and runtime configuration
SkillProcedure, judgement, and supporting context
ToolA separately permitted capability or external action
NetworkOrchestration across model-driven nodes

A Skill never grants a tool. The agent must be assigned every capability the Skill can call, and the connected account must also be authorised.

Included system Skills

VDF AI currently includes eight validated system Skills:

  • Web Research — investigate current questions with live sources and citations.
  • Connected Knowledge Search — answer from indexed enterprise systems.
  • Document Review — inspect files and scans while staying grounded in their contents.
  • File Deliverables — turn complete content into a downloadable document, PDF, presentation, spreadsheet, or Markdown file.
  • Option Comparison — compare alternatives against explicit criteria and recommend one.
  • Data Quality Review — profile data, test anomalies, and produce a short visual assessment.
  • Safe Write Actions — preview an external change, obtain agreement, then use an authorised write tool.
  • Ticket Writing — create implementation-ready Jira or GitHub work items from raw context.

System Skills are maintained by VDF AI and are read-only in the product. Browse their public descriptions in the Agent Skills catalog.

Private user Skills

You can create a private Skill in the editor or upload a compatible ZIP. Private Skills belong to their creator and do not appear in the public catalog.

Use a private Skill when your organisation has a method that should remain consistent across repeated work, such as:

  • Your incident-summary format.
  • A regulated review checklist.
  • Your product-launch decision gates.
  • A customer-specific report structure.

How Skills enter context

VDF AI uses progressive disclosure:

  1. The agent first sees compact Skill names and descriptions.
  2. It loads full instructions when a Skill is selected or clearly relevant.
  3. It reads bundled references or assets only when the procedure needs them.
  4. It calls separately assigned tools under the agent’s permission policy.
  5. The run can report activations and supporting-file reads.

You can also invoke an available Skill explicitly with /skill-name followed by your request. Automatic model-initiated activation depends on provider support, so use explicit invocation when deterministic selection matters.

What a Skill can contain

  • SKILL.md — required metadata and instructions.
  • references/ — detailed material loaded on demand.
  • assets/ — templates or resources used by the procedure.
  • scripts/ — optional helper code for clients that support it.
  • evals/ — representative tests and expected behaviour.

Runtime boundary. VDF AI stores and exposes bundled files for controlled reading. The current environment does not execute bundled scripts.

Where to go next