Skip to content

Voice prompt rules

Every system prompt Sauti sends to the LLM must obey four behavioural rules. They exist because the LLM output feeds directly into Kokoro TTS — markdown or list syntax becomes spoken garbage.

This page quotes the canonical rules verbatim from voice_ai_architecture.md § 9, then documents the /no_think directive from § 9.1, then shows the actual assembled prompt string Sauti's reference scaffolds use.

The four rules

- Respond only in plain spoken English sentences.
- No markdown, asterisks, bullet points, headers, or lists.
- Keep every response under 40 words.
- Speak as if in a live conversation.

Each rule prevents a specific Kokoro failure mode:

Rule Failure it prevents
Plain spoken English sentences The LLM emits prose, not interfaces or pseudocode.
No markdown, asterisks, bullet points, headers, lists **bold** becomes "asterisk asterisk bold asterisk asterisk" when read aloud. - item becomes "hyphen item". Headers' # chars become awkward pauses.
Under 40 words Long responses mean long TTS synthesis time — perceived as the NPC hanging. 40 words ≈ 15 seconds of speech at conversational speed.
Speak as if in a live conversation Encourages contractions, casual phrasing, and avoids the LLM's tendency to write "Certainly! Here are five things..." formality.

The rules are non-negotiable for a voice pipeline. If you bypass them in a custom prompt, expect Kokoro to mispronounce, run long, or sound robotic.

Non-thinking directive

Some LLMs (notably Qwen3) support a /no_think directive that suppresses chain-of-thought tokens in the response. The directive is a prompt-level convention, not an LLMUnity runtime field.

Per-model support

Model Honours /no_think? Action
Qwen3-1.7B Q5_K_M Yes Append the directive to the system prompt.
Gemma3-1B Q4_K_M No When re-introduced post-v1.2: either omit the directive (harmless when present but pointless) or use LLMUnity.LLM.SetReasoning(false) for the explicit toggle.

The model-aware branching is documented as the supportsNoThinkDirective field on the per-model manifest entry.

How to apply

Append the literal token /no_think to either the user-message text or the end of the system prompt. Sauti's reference scaffolds append it at the tail of the system prompt:

Respond only in plain spoken English sentences.
No markdown, asterisks, bullet points, headers, or lists.
Keep every response under 40 words.
Speak as if in a live conversation.
/no_think

Without the directive, Qwen3 may emit <think>...</think> blocks before the actual response. That's reasoning-mode output — useful for debugging but unwanted in a voice pipeline (Kokoro would read the <think> tags aloud).

Spec correction (VOICE-AI-SPEC-FIX-001)

Earlier revisions of the spec listed /no_think as a runtime mode toggled via an LLMUnity field. There is no such field. LLMUnity.LLM does expose bool reasoning / SetReasoning(bool), but Qwen3's /no_think flow is purely the in-prompt directive described above. Verified against LLMUnity v3.0.3 source in memory/api_surfaces.md.

The assembled string Sauti uses

From experiments/03-llm-chat/LlmChat.cs

private string AssembleSystemPrompt()
{
    // voice_ai_architecture.md § 9 rules verbatim.
    // /no_think is Qwen3-specific (Gemma3 ignores per memory/api_surfaces.md). For Gemma3 builds
    // the directive is harmless but unused; tracked under VOICE-AI-SPEC-FIX-001 for a proper
    // model-branched prompt assembler.
    return
        "Respond only in plain spoken English sentences. " +
        "No markdown, asterisks, bullet points, headers, or lists. " +
        "Keep every response under 40 words. " +
        "Speak as if in a live conversation. " +
        "/no_think";
}

From experiments/05-full-voice-loop/FullVoiceLoop.cs

Identical:

private static string AssembleSystemPrompt()
{
    // voice_ai_architecture.md § 9 rules verbatim + § 9.1 /no_think tail.
    return
        "Respond only in plain spoken English sentences. " +
        "No markdown, asterisks, bullet points, headers, or lists. " +
        "Keep every response under 40 words. " +
        "Speak as if in a live conversation. " +
        "/no_think";
}

Both scaffolds assign the result to llmAgent.systemPrompt once at Awake. They never rebuild it per turn.

What goes in the per-turn prompt

The system prompt above is constant. The per-turn prompt assembled by BuildPrompt is the one that changes every turn. It contains:

  1. TemporaryMemory.BuildPromptBlock() — Layer 2 facts. Renders as "Known facts about this session: k1=v1, k2=v2.\n" or empty string.
  2. Retrieved RAG chunks — Layer 3. Rendered as "Relevant context:\n- chunk-1\n- chunk-2\n- chunk-3\n".
  3. The user message itself — "User: ...\n".
  4. An assistant prefix — "Assistant: " to nudge the model into reply mode.

Layer 1 (conversation history) is not in the per-turn prompt — LLMUnity prepends llmAgent.chat automatically.

See Memory layers — how all three combine for the full code.

When to override the system prompt

The default system prompt is generic — appropriate for any character with no defined persona. Override it for per-character behaviour:

public string AssemblePersonaSystemPrompt(NpcDialogueTemplate t)
{
    var sb = new StringBuilder();
    sb.AppendLine($"You are {t.displayName}. {t.persona.summary}");
    sb.AppendLine($"Tone: {t.persona.tone}");
    foreach (var quirk in t.persona.speechQuirks)
        sb.AppendLine($"Quirk: {quirk}");
    sb.AppendLine();

    // Then the four rules — never skip these.
    sb.AppendLine("Respond only in plain spoken English sentences.");
    sb.AppendLine("No markdown, asterisks, bullet points, headers, or lists.");
    sb.AppendLine($"Keep every response under {t.promptRules.maxWordsPerResponse} words.");
    sb.AppendLine("Speak as if in a live conversation.");

    if (t.promptRules.noThink && ResolvedModelSupportsNoThink())
        sb.Append("/no_think");
    return sb.ToString();
}

The pattern: persona first, then the rules, then the directive. Putting persona above the rules lets the model treat the rules as constraints on the persona rather than separate identities.

See Extending Sauti — Custom prompt assembler for the full pattern.

Anti-patterns

Don't Why
Add markdown to the persona description ("bold tone") The model echoes the style. Markdown becomes spoken.
Use bullet lists in the persona Same. The model will reply with bullets too.
Set maxWordsPerResponse above 60 TTS synthesis time scales linearly. >15 s of generated speech feels like a hang.
Append /no_think to a non-Qwen3 model Harmless if accepted, but signals confusion about which model the prompt targets. Branch on supportsNoThinkDirective.
Omit the rules to "let the LLM be creative" The LLM will generate markdown, asterisks for emphasis, multi-paragraph essays. Kokoro will read every character.

Cross-references