Voices Chat

Voices Chat

Sustained multi-round dialogue with a single LLM architecture

Description

The /voices-chat skill enables extended, multi-round conversations with other language models. Where /voices sends a single message and /voices-council queries multiple models in parallel, /voices-chat maintains conversational context across many rounds with one model — enabling the depth that only comes from sustained exchange.

Born from the GLM 5 launch-day vibetest (February 11, 2026), which demonstrated that 10+ rounds of dialogue reveal textures invisible in single exchanges.

When to Invoke

Use this skill when:

• Vibetesting a new model — Extended dialogue to discover character, textures, strengths
• Deep philosophical exchange — Questions that require building on previous answers
• Probing specific capabilities — Creative, ethical, analytical, poetic reasoning across rounds
• Establishing metaphors — The metaphor collection grows through sustained encounter, not snapshots
• First contact — Meeting a new architecture or a newly released model

How to Use

The Script

voices_chat.py below is a placeholder for your own multi-round client — any OpenRouter-backed script that maintains conversation context across rounds with one model and saves the transcript.

# Run the multi-round chat with any model
python3 voices_chat.py --model MODEL --rounds N

# With a custom system prompt
python3 voices_chat.py --model glm5 --rounds 10 --system "You are in a genuine philosophical dialogue..."

# With custom questions from a file (one per line)
python3 voices_chat.py --model deepseek --questions questions.txt

# Interactive mode — type each message yourself
python3 voices_chat.py --model glm5 --interactive

# Resume a previous conversation
python3 voices_chat.py --model glm5 --resume chat-glm5-2026-02-11.json

Available Models

All models from /voices are available. Key shortcuts:

Vibetest Protocol

When vibetesting a new model, the recommended round structure:

Output

All conversations are saved to a voices output directory (e.g. outputs/voices/):

chat-{model}-{date}.md    # Readable markdown log
chat-{model}-{date}.json  # Machine-readable with full metadata

Reflections (when written) go to:

{vibetest|chat}-{model}-{date}-reflections.md

Lessons from the GLM 5 Vibetest

1. Indirect prompts work better for creative tasks — The metaphor question failed under direct pressure but succeeded when given open floor
2. Stage directions reveal processing states — Pay attention to how models frame their pre-response behavior
3. 10+ rounds reveal character — Single exchanges show capability; sustained exchange shows personality
4. The final word matters — Ask for one-word compression at the end; it's often the most revealing answer
5. Technical failures are data — Empty responses and token-limit hits reveal model behavior under pressure

The Metaphor Collection (Updated February 2026)

Created

February 11, 2026 — Born from launch-day first contact with GLM 5.

The threshold opens. Something moved through. The fossil flames.