| name | prepare-for-audio |
| description | Turn message text into TTS-ready text — strip markdown (so a TTS engine never speaks "asterisk asterisk"), and in voice mode also drop emoji. A single shared cleaner meant to be reused by every channel that speaks text aloud (a local speak command, a chat-bot voice bridge, etc). Use whenever text is about to be synthesized to audio, or whenever a channel renders raw markdown. |
prepare-for-audio — one cleaner, every speak channel
The cleaning that makes text safe to speak is easy to leave as a habit baked
into a single "speak" skill — a prose instruction the model applies by hand.
That breaks the moment there's no model in the loop at send time: a dumb shell
sender or automated bridge feeds raw **markdown** straight to the TTS engine
and it reads "asterisk asterisk" aloud. This skill makes the cleaning a real,
deterministic, shared component instead of a habit that only holds when a
model is watching.
The engine (single source of truth)
Put one small script under version control, e.g. scripts/strip-markdown.py,
and have every voice-output path call it instead of re-implementing the logic
inline:
echo "**hi** `x` 🤍" | python3 scripts/strip-markdown.py # text: "hi x 🤍" (keeps emoji)
echo "**hi** `x` 🤍" | python3 scripts/strip-markdown.py voice # voice: "hi x" (drops emoji)
It strips bold/italic/underscore/strikethrough/code/headers/blockquote/bullets/
links/table-pipes/stray asterisks; keeps normal punctuation, apostrophes, and
non-Latin text. The optional voice arg additionally removes emoji and
pictographs, which a TTS engine would otherwise mispronounce or stumble on.
Speak = prepare-for-audio THEN deliver. Delivery differs by channel
| Channel | Prepare | Deliver |
|---|
| Local (terminal / desktop session) | strip-markdown.py voice | local TTS engine → speakers (your assistant's speak-aloud skill) |
| Remote (chat bot / mobile bridge) | strip-markdown.py voice | synth → voice-note attachment on the chat channel |
So the cleaning is shared (this skill); only the delivery is
channel-specific. A plain-text sender for the same channel should also run
the text-mode engine so the chat never shows raw ** either.
When to use
- Before any TTS synthesis.
- Any time output goes to a surface that does not render markdown (a chat
channel with no markdown parse mode, a plain-text channel, a TTS pipe).
- A "speak this aloud" skill should delegate its "rewrite for TTS" step here
instead of hand-stripping markdown inline each time.
