Skip to content
Back to skills

Ac Tools Voice User

ASecurity

Use the say tool for short one-way spoken alerts to the user. Use when you finish a meaningful task, need attention, confirmation, approval, or input, or when the user explicitly asks for voice output. Keep spoken text brief and put the full details in writing.

  • 30 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added September 6, 2026
ai-agents

Security analysis

A100/100

Scanned September 6, 2026

npx -y skills add MatiasComercio/agentic-config --skill ac-tools-voice-user --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Ac Tools Voice User?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Ac Tools Voice User
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/matiascomercio-ac-tools-voice-user/badge)](https://www.skillsdirectory.com/skills/matiascomercio-ac-tools-voice-user)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: ac-tools-voice-user
description: Use the say tool for short one-way spoken alerts to the user. Use when you finish a meaningful task, need attention, confirmation, approval, or input, or when the user explicitly asks for voice output. Keep spoken text brief and put the full details in writing.
project-agnostic: true
allowed-tools:
  - say
---

# Voice User

Use the `say` tool for concise one-way spoken communication with the user.

## When to use

- You finished a meaningful task and want to alert the user.
- You need the user's attention, confirmation, approval, choice, or missing input.
- The user explicitly asks for spoken output.
- Only in the top-level agent runtime. Subagents cannot use `say`.

## How to use

- Call the `say` tool with a short spoken message.
- Prefer under 50 words when asking for attention or confirmation.
- Use this shape when possible:
  - `<project> - <what you need>`
- After the spoken alert, give the full explanation in the normal written response.

## Keep spoken output short

Do not speak:

- long explanations
- code
- secrets, tokens, or credentials
- stack traces
- large diffs or long file paths

## Tool shape

```json
{
  "text": "play - say extension updated; please reload.",
  "voice": "Samantha",
  "rate": 190
}
```

Only `text` is required. If `voice` or `rate` is omitted, the extension may use the saved defaults.

## Notes

- The extension may configure automatic voice as `off`, `always`, or `long`.
- Even when automatic voice is off, use the `say` tool when the user explicitly wants spoken output.
- Subagents do not have access to `say`. If a subagent finishes work or needs attention, it should report that in writing and let the parent agent decide whether to speak.
- Long explanations stay in writing; voice is only the concise alert.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.