say-it

name	say-it
description	Speak content aloud via Kokoro neural TTS (local, offline). Use when the user says "read it for me", "read it to me", "play it in audio", "say it", "speak it", "read that aloud", "/say-it", or asks to hear something spoken instead of reading it.
user_invocable	true
args	Optional. What to speak — "last response", a section name, a file path, or inline text.
allowed-tools	Write, Bash(kokoro-say *), Bash(pkill *), Bash(kill *), Bash(ls *)

Say It — Speak Content Aloud

Prerequisites

Requires the voice stack installed via /setup-voice:

• kokoro-say in ~/.local/bin/
• Kokoro models at ~/.cache/kokoro/

Run /setup-voice first if kokoro-say is not found.

Model Selection

• Default model: Haiku 4.5 — markdown strip + kokoro-say invocation is mechanical
• Promote to Sonnet when: user asks for a condensed or rephrased spoken version ("say it like a summary", "read me the gist")

Triggers

• "read it for me", "read it to me", "read that aloud", "read the response aloud"
• "play it in audio", "play this in audio"
• "say it", "say that", "speak it", "speak that"
• "/say-it"
• Any "can you speak / read / say" request aimed at a response or file

What to speak

• Default: the most recent substantive response you produced (not the reply you are about to give).
• User specifies a section/file: speak only that part.
• User provides text inline: speak it verbatim (or rephrase only if they asked).

If the content is over ~5000 words, confirm first.

Voice selection

Context	Voice flag
Default (tech/Claude Code sessions)	-v af_bella
If user says "warm" / "Heart"	-v af_heart
If user says "male" / "Adam"	-v am_adam
If user says "Michael"	-v am_michael
If user says "British" / "Emma"	-v bf_emma

Full voice list: af_heart, af_bella, af_sky, bf_emma, bf_isabella, am_adam, am_michael

The steps

1. Rewrite for TTS. kokoro-say reads text literally. Before invoking:

   • Strip #, *, _, backticks, table pipes, bullet characters.
   • Replace code blocks with a short spoken description or skip.
   • Expand obvious initialisms on first use ("PR" → "pull request", "TTS" → "text to speech").
   • Break long paragraphs into short sentences — periods drive pacing.
   • Spell tricky filenames phonetically (CLAUDE.md → "claude markdown").
   • Keep natural commas and periods; the voice uses them for rhythm.

2. Invoke in the background:

   kokoro-say -v af_bella "cleaned text here" &
   

   Always background with & — do not block the session.

3. Tell the user it's playing and how to stop it.

Stopping

Any of: "stop", "quiet", "cancel", "enough", "shush" →

pkill -f "kokoro-say"; pkill afplay

Don'ts

• Never use macOS say — always use kokoro-say.
• Don't read raw markdown aloud — strip it first.
• Don't block — always background.
• Don't speak content containing secrets or credentials without confirming.
• Don't re-speak the same content twice without being asked.