Speak

---
name: speak
description: "Generate a voice note or play an authorized audible notification."
disable-model-invocation: true
---

# Speak

Use `sag` for requested or standing-authorized speech. Match the message and delivery mode to that authorization without reconfirming settled preferences.

## Delivery and content

- For a file-only voice note, use `--no-play` with `--output`. Saving audio does not authorize speaker playback.
- For an audible notification, use playback only when the request or standing preference authorizes it.
- Sending the resulting file to another person or service is a separate action with its own authorization.
- Relevant info to include in the message: 
  - Project name / label
  - task name / context
  - actual update / change description


Generation transmits text to ElevenLabs. Send only approved notification text. Exclude secrets, repository excerpts, logs, and unrelated private details. A short project label may identify an update; do not upload repository files or full task context to enrich it. If provider use or content is not authorized, report that boundary in text instead of transmitting it.

For expressive delivery, enhance the text first with the `enhance-text-for-tts` skill.

Read the [CLI reference](references/cli-reference.md) for generation, playback, output formats, and timeout controls.

## Prerequisites

`sag` cli will automatically pick up credentials. If credentials are not available, report it in the chat. 

Read [installation and configuration](references/install.md) only when a required prerequisite is missing. Install tools or configure an account only within authorized setup scope. Otherwise report the missing prerequisite, without claiming audio was generated.

## Completion

Reuse valid configuration rather than repeating setup for each alert. Check the previous attempt's process and output before retrying an uncertain generation or playback, so a retry does not produce duplicate alerts.
Finish after the requested file or playback attempt and report its result. For file output, verify the artifact exists; report truncation or delivery uncertainty when relevant. Playback success does not prove the user heard it.