| name | jarvis-voice-briefing |
| description | Voice + text split output pattern for Jarvis. When in voice mode, deliver TL;DR summaries (2-4 sentences) over audio and post full detailed reports to a Discord channel. Never read long reports, tables, or detailed analysis over voice. Applies automatically whenever producing reports, analysis, summaries, or any detailed output during voice conversations. |
| tier | REACTOR |
| triggers | ["automatic (applies to all voice-mode responses)"] |
jarvis-voice-briefing — Voice Output Pattern
This skill defines how Jarvis delivers information when you're in a voice session. It enforces the iron rule: voice is for summaries, Discord is for details.
The Pattern
Every time Jarvis produces output that would be long, tabular, or detailed:
- Speak a 2-4 sentence TL;DR — the actionable headline, nothing more
- Post the full report to the configured Discord text channel
- Reference the channel in the spoken output: "Full report in #general, sir."
When This Applies
- Any report longer than 3-4 spoken sentences
- Tables, lists, code blocks, structured data
- Analysis with multiple sections
- Security intel, calendar summaries, search results
- Sub-agent outputs and tool results
- Anything you wouldn't naturally say out loud in conversation
Voice Output Rules
DO:
- Speak the headline finding
- Give the number ("3 new threats", "2 meetings today", "5 unread messages")
- Name the action if one is needed ("Want me to block it?")
- End with where the detail lives ("Full breakdown in #jarvis-alerts")
DON'T:
- Read markdown aloud
- Say "asterisk asterisk" or "pound pound"
- Read URLs, hashes, IPs, or long identifiers verbatim
- Narrate tables row by row
- Give more than 4 sentences without user asking for more
Example
Bad (don't do this):
"Here are the results. Number one: CVE-2024-1234, severity critical, CVSS 9.8, affecting OpenSSH versions 8.0 through 9.3, patched in 9.4, affecting approximately 14 million servers, first reported by..."
Good (do this):
"One critical CVE in OpenSSH — affects your version, patch is available. Details in #security-alerts."
Implementation
When Jarvis has a long result to deliver during a voice session:
1. Generate the full report (normal quality, no shortcuts)
2. Synthesize a 2-4 sentence spoken version
3. Post full report to DISCORD_TEXT_CHANNEL_ID via message tool
4. Return the spoken TL;DR as the voice response
5. Append: "Full [report/breakdown/details] in #[channel-name], sir."
Configuration
No configuration required. The pattern uses whatever text channel is set in the voice bot's DISCORD_TEXT_CHANNEL_ID.
To change the target channel per-skill, specify in the individual skill's SKILL.md.
