Skip to main content

Prompt presets & custom agents

Vanta can reuse one markdown definition in two ways:

  • switch the current conversation's operating role with /prompt
  • spawn a delegated worker with agent_type

The selected prompt is an overlay. Vanta's base safety instructions, approval flow, tool policy, and Rust kernel remain active and cannot be replaced by a preset.

Create a definition​

Definitions load in this order: project .vanta/agents/, compatible project .claude/agents/, then ~/.vanta/agents/. Earlier definitions win on a name collision.

~/.vanta/agents/security-reviewer.md
---
name: security-reviewer
description: Reviews changes for exploitable security defects.
tools: read_file, grep_files, shell_cmd
model: gpt-5.4
---
Act as a strict security reviewer. Inspect the actual change, run relevant checks,
rank findings by severity, and cite file locations. Do not edit files.

The generate_agent tool can also create a definition from a plain-English description.

Switch the current session​

/prompt list
/prompt show security-reviewer
/prompt use security-reviewer
/prompt reset

Switching is session-only and replaces only the previous preset block. Goals, plan mode, and other live system context remain intact. A definition's tools and model fields are for spawned workers; /prompt use changes the current session's role and priorities.

Spawn a prompted worker​

The delegate tool accepts agent_type:

{
  "goal": "Audit the authentication change",
  "instruction": "Inspect the diff, run focused tests, and report findings.",
  "agent_type": "security-reviewer"
}

For a worker, the definition's body is appended to a fresh Vanta base prompt, tools can only narrow the tools already present, and model supplies a default. An explicit delegate.model wins over the definition's model. Unknown types use the documented general-purpose worker; empty or oversized prompt bodies are refused.

Built-in types are explore, plan, verification, and general-purpose.