Configure Verbosity

Verbosity controls how chatty the agent is — both its pre-tool acknowledgement ("hmm, let me check") and its post-tool spoken reply (fact + optional follow-up). The LLM emits the preamble inline in its response stream, not via a separate parallel call. See Natural-Sounding Fillers for the design.

Levels

Each level rewrites two blocks in the system prompt — preamble (while a tool runs) and post-tool response (the spoken answer):

Level	Preamble	Post-tool response
silent	(none)	10-15 words; top fact only, no follow-up offer
brief (default)	2-4 words — "one sec"	12-18 words; fact + "want more?"
narrated	4-8 words — "okay, let me look that up"	18-25 words; fact + one supporting detail + follow-up
chatty	up to 12 words	25-40 words; fact + two details + warm follow-up

Post-tool response sizing is backed by CHI 2025 (Kim et al., "Listening vs Reading: Information Density in Voice UIs") — optimal spoken summary is 18-25 words, and past 40 words users barge-in or skip 3× more often.

The model writes its own preamble + answer each time, conditioned on the user's actual message. There's no phrase pool. Examples in the prompt are explicitly marked "do NOT copy verbatim."

Latency tiers gate progress narration

Even with verbosity set, FAST and MEDIUM tools just rely on the inline preamble — no separate progress channel. Only SLOW tools (3s+) get a periodic generated "still working" line:

Tier	What you hear
FAST (calculator, get_datetime)	Inline preamble + answer in one shot
MEDIUM (web_search, deep_research, a2a_dispatch)	Inline preamble + answer
SLOW (slow_research)	Inline preamble + periodic generated progress + final delivery

Pick at startup

VERBOSITY=narrated docker compose up -d protovoice

Change between sessions

The dropdown rewrites the prompt for future connections:

curl -X POST http://localhost:7867/api/verbosity \
  -H "Content-Type: application/json" \
  -d '{"level":"chatty"}'

Read the current value:

curl http://localhost:7867/api/verbosity
# → {"verbosity":"brief"}

Why not mid-session: the verbosity setting is baked into the system prompt at connect time. Changing it during a live session would require rewriting the live LLM context, which we deliberately don't do (consistent with how skill switching works — see Personas & Skills).

Backend-aware prosody

The TOOL USE prompt block also includes a backend-specific style hint:

• Fish Audio S2-Pro — the LLM is told it may use [softly], [pause], [hmm], [thinking], [whisper] tags sparingly. Sample output: [softly] hmm, let me check.
• Kokoro 82M — the LLM is told NO bracketed tags (Kokoro speaks them as literal sounds). Sample output: okay, let me check.

The active skill's tts_backend decides which style hint gets injected. So a Fish-backed persona naturally speaks with prosody tags; a Kokoro-backed persona speaks plain.

Tuning progress cadence (SLOW tools)

Two-tier cadence matching Alexa's production pattern + the arXiv 2507.22352 finding that >4 s of unfilled silence degrades perceived QoE:

• progress_first_secs — wait this long after the tool starts before the first "still working" line. Default 2.0 s.
• progress_second_secs — wait this long again before emitting a second ack (~8 s total wall-clock). Default 6.0 s.

After the second ack the loop stops — narrating past ~8 s reads as performative. A SLOW tool that finishes inside progress_first_secs only gets the LLM's inline preamble, no progress narration.

Persona override

Skill YAMLs can override session verbosity per persona:

slug: chef
filler_verbosity: brief   # Chef Bruno is terse — he's busy

The override applies at the next connect; the verbosity-shaped prompt block bakes into that session's system message.