kitt:speak -- SPEAK Marker Reference

How to embed speech markers in Claude Code responses. Loaded by /kitt:speak-enable, not invoked directly.

Marker Format

<!-- SPEAK -->**brief summary here**                              (defaults: session voice, 180 wpm)
<!-- SPEAK(voice:Samantha) -->**with custom voice**               (custom voice)
<!-- SPEAK(rate:250) -->**with custom speed**                     (custom speed)
<!-- SPEAK(voice:Samantha rate:250) -->**both**                   (both)

Defaults: your session voice (set at enable time), rate 180 wpm.

Control Markers

These are emitted ONLY by /kitt:speak-enable and /kitt:speak-disable, never in regular responses:

<!-- SPEAK-ENABLE(voice:Karen) -->**Speech enabled.**             (enable session speech)
<!-- SPEAK-DISABLE -->**Speech disabled.**                        (disable session speech)

The hook uses these to create/remove per-session enabled files and voice locks.

When to Include Markers

DO include a <!-- SPEAK --> marker for:

• Task completions ("Done, tests pass.")
• Errors or failures ("Build failed.")
• Questions to the user ("Which approach?")
• Key findings or status updates ("Found the bug.")

DO NOT include markers for:

• Code blocks, file listings, verbose output
• Intermediate tool-call-only responses
• Long explanations (pick a one-line headline instead)
• Flattery or low-value acknowledgment ("good call", "will do", "got it", "sounds good", "absolutely")

Speech Content Rules

Speech must be informative, not agreeable. Describe WHAT you're doing, not THAT you agree.

Bad (blocked by hook)	Good
"Good call, doing that."	"Adding flattery filter."
"Sure, will do."	"Updating the hook script."
"Got it, on it."	"Refactoring the parser."
"Absolutely, great idea."	"Building the new endpoint."

Speech must be plain English words only. No hashes, hex codes, line numbers, or alphanumeric noise. Humans cannot follow spoken gibberish like commit SHAs or UUIDs.

Bad	Good
"Pushed d651e91."	"Pushed to main."
"Fixed line 133."	"Fixed the shell config."
"Error 0x8BADF00D."	"Memory error found."
"Version 3.1.4-rc2."	"New version pushed."

Short version labels like "v3" are OK.

Rules

• Keep spoken text under 5 words -- a tiny headline
• Plain English words only -- no hashes, hex, line numbers, or codes
• One marker per response is usually enough
• The marker is invisible to the user (HTML comment) but the bold text renders nicely
• Multiple markers in one response play sequentially

Available Voices

Voice is chosen at enable time. First unclaimed voice from the pool (Kokoro TTS via mlx-audio):

Voice	Style	Grade
af_heart	American Female	A
af_bella	American Female	A-
bf_emma	British Female	B-
af_nicole	American Female	B-
am_fenrir	American Male	C+
am_michael	American Male	C+
am_puck	American Male	C+
bm_george	British Male	C
bm_fable	British Male	C
af_aoede	American Female	C+
af_kore	American Female	C+
af_sarah	American Female	C+

Override per-marker with <!-- SPEAK(voice:af_bella) -->. Full voice list at hexgrad/Kokoro-82M on HuggingFace.