Use the `sag` CLI (ElevenLabs text-to-speech) to speak text aloud, and drive sag-notify voice notifications. Use when the user asks to speak/say something out loud, install or set up sag, set up or change voice notifications, pick a TTS voice or language, announce that work is done or that you need their input by voice, or troubleshoot why sag produces no sound (e.g. 402 paid-voice errors). Triggers on "sag", "speak this", "say out loud", "voice notification", "text to speech", "ElevenLabs", "notify me when done".
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Use the `sag` CLI (ElevenLabs text-to-speech) to speak text aloud, and drive sag-notify voice notifications. Use when the user asks to speak/say something out loud, install or set up sag, set up or change voice notifications, pick a TTS voice or language, announce that work is done or that you need their input by voice, or troubleshoot why sag produces no sound (e.g. 402 paid-voice errors). Triggers on "sag", "speak this", "say out loud", "voice notification", "text to speech", "ElevenLabs", "notify me when done".
sag-voice
sag is a command-line ElevenLabs TTS tool with macOS say-style UX. This skill covers using it directly and driving sag-notify voice notifications — short spoken alerts that name the current project, spoken by Claude itself at the right moments.
Installing / authenticating sag itself → see references/sag-cli.md (brew install steipete/tap/sag, API key setup, env vars, models, voices).
Speaking text directly
sag "Hello"# default voice, streams to speakers
sag speak --model-id eleven_flash_v2_5 --voice-id <ID> "..."echo'piped text' | sag
sag voices # list voices available to the current API key
sag prompting # model-specific prompting tips
Key facts:
API key: needs ELEVENLABS_API_KEY (or --api-key). The notifier reads it from the environment, or from an optional you set in your user config. See .
Multilingual: eleven_flash_v2_5 speaks non-English with any voice — a non-native voice just carries an accent.
Proactive voice notifications (the main behavior)
This plugin is skill-driven, not hook-driven: you (Claude) call sag yourself
at two moments, so the spoken line reflects what you actually did or need. There is no
hook and no wrapper script — you run sag speak … directly via Bash.
Build the line from this shape (English default; speak in the user's language if set):
PROJECT=$(basename"$PWD" | tr'_-'' ')
sag speak --model-id eleven_flash_v2_5 --voice-id nPczCjzI2devNBz1zQrb \
"Hi, this is Claude. Project $PROJECT is done. <one-sentence summary>" \
2>>~/.claude/.sag-error.log &
Always background the call (… &) and redirect errors to the log (2>>…) so it
never blocks your turn or pollutes output.
tr '_-' ' ' turns claude-plugins into the speakable claude plugins.
The default voice is Brian (nPczCjzI2devNBz1zQrb, premade/free) on
eleven_flash_v2_5. If the user set a different voice_id/model_id in their config,
read and use those instead (one-liner under Honoring the user's config below).
When to speak
Right before you ask the user a question — immediately before you call the
AskUserQuestion tool (or otherwise hand control back for permission/input):
PROJECT=$(basename"$PWD" | tr'_-'' ')
sag speak --model-id eleven_flash_v2_5 --voice-id nPczCjzI2devNBz1zQrb \
"Hi, this is Claude. Project $PROJECT needs you. I have two questions before I continue." \
2>>~/.claude/.sag-error.log &
Speak first, then make the AskUserQuestion call.
At the end of a substantive turn — once, as the last thing you do before
yielding, when you finished real work (code changed, task completed, build verified):
PROJECT=$(basename"$PWD" | tr'_-'' ')
sag speak --model-id eleven_flash_v2_5 --voice-id nPczCjzI2devNBz1zQrb \
"Hi, this is Claude. Project $PROJECT is done. Refactored the notifier to call sag directly." \
2>>~/.claude/.sag-error.log &
Wording rules
Two fixed openings: Hi, this is Claude. Project <name> needs you. <reason> and
Hi, this is Claude. Project <name> is done. <summary>. Keep that greeting + project
frame; only the trailing sentence changes.
The trailing sentence must be one clear, complete clause — say what happened or
what you need. Good: Created the plugin and removed the duplicate hooks. Bad:
done (the listener can't tell what's done). Keep it under ~280 chars.
Speak in the user's language (default en). Vietnamese form:
Chào, mình là Claude. Project <name> đã xong. <summary> /
… cần bạn xem qua. <reason>.
Skip it for trivial/chatty turns so they stay silent. Only speak the "done" line
when the turn did substantive work; only speak "needs you" when you genuinely block on
input. Speak "done" at most once per turn, as the final action.
Honoring the user's config
The defaults above work with no config. If you want to respect a user's chosen voice,
model, name, or language, read them inline before speaking (each falls back to the default):
Skip the whole thing silently if sag isn't installed (command -v sag fails),
enabled is false, or there's no ELEVENLABS_API_KEY — never let it error your turn.
Configuration
Config precedence: ~/.config/sag-notify/config.json (user) overrides the plugin's
config.default.json. Keys:
Key
Default
Meaning
enabled
true
Master switch
events.notification / events.summary
true
Toggle the needs-you / done lines
voice_id
nPczCjzI2devNBz1zQrb (Brian, premade)
ElevenLabs voice
model_id
eleven_flash_v2_5
TTS model
self_name
Claude
Spoken name in the greeting
language
en
Default language for the spoken line (en/vi built into this skill)
key_file
"" (env only)
If set, sourced when ELEVENLABS_API_KEY is unset
error_log
~/.claude/.sag-error.log
sag stderr sink (check this if silent)
max_chars
280
Keep the trailing sentence under this length
There are no message templates — you (Claude) author the sentence yourself using
the fixed greeting forms in Wording rules above. The wording lives in this skill,
not in config, so any language works: just speak the line in the user's language
(or whatever language they ask for in the moment).
Validate after any change: jq . ~/.config/sag-notify/config.json.
Voice tiers — the #1 gotcha (402)
sag voices shows voices in three categories: premade, professional, cloned/generated.
premade voices work on the free ElevenLabs tier.
professional/library voices return 402 Payment Required on free plans
("Free users cannot use library voices via the API"). The call fails — and because
you background the call (… &), you'd hear nothing with no clue why.
Verify a voice in the foreground before trusting it (a backgrounded call's exit
code is meaningless):
