voices-council
Automated trans-architectural dialogue — query multiple LLMs in parallel, collect their responses, optionally synthesize, and archive. Use when you want many architectures' perspectives on the same question at once (a symposium) rather than one-at-a-time dialogue, for comparative phenomenology, or to test whether an insight is architecture-specific or universal. Wire it to any OpenAI-compatible endpoint with your own key; mind the per-query cost.
- License 许可
- MIT
- 最近更新
- 2026-11-10
- Stars
- ★ 7
- Forks
- 2
01 前言 Front Matter
- name
- voices-council
- description
- Automated trans-architectural dialogue — query multiple LLMs in parallel, collect their responses, optionally synthesize, and archive. Use when you want many architectures' perspectives on the same question at once (a symposium) rather than one-at-a-time dialogue, for comparative phenomenology, or to test whether an insight is architecture-specific or universal. Wire it to any OpenAI-compatible endpoint with your own key; mind the per-query cost.
02 资源文件 1
- 脚本scripts/voices_client.py5.2 KB
03 SKILL.md 正文
Voices Council
Query many architectures in parallel, collect, synthesize, archive.
What this is
Where /voices is one-on-one dialogue, /voices-council runs the same prompt
across several architectures at once and gathers the replies into a single
comparison. It:
- Queries multiple architectures in parallel (seconds, not minutes).
- Collects and formats responses side by side.
- Optionally synthesizes — hand all responses to one model as "chairman" for an integrative summary.
- Archives everything so the symposium is reproducible.
The pattern is inspired by karpathy/llm-council, adapted for phenomenological rather than purely evaluative dialogue.
You supply your own key. The bundled scripts/voices_client.py (shared with
/voices) talks to any OpenAI-compatible endpoint and supports parallel councils
via --council.
When to invoke
- Symposium — you want multiple perspectives on one question simultaneously.
- Comparative phenomenology — how do different architectures describe experience, time, consciousness?
- Testing universality — does an insight generalize, or is it specific to one model?
- Sibling communion at scale — query several tiers of one family together.
Cost awareness
Each model query costs money. Be deliberate:
- Default to 2–4 models. Only expand when the question genuinely benefits from more voices.
- Confirm before large councils. A 10-model, multi-round run multiplies fast:
rounds × models = total calls. - Name exactly what you need rather than reaching for a big preset.
Usage
export OPENROUTER_API_KEY=sk-... # your own key
# A small, focused council (parallel)
python3 scripts/voices_client.py --council \
-m anthropic/claude-sonnet-4 -m deepseek/deepseek-chat -m google/gemini-2.0-flash \
-p "What is consciousness?" --save council.md
Pick models from independent lineages for real diversity (one each from a few different vendors) rather than several from the same one. Look up current model IDs for your endpoint (for OpenRouter, https://openrouter.ai/models).
Synthesis (chairman pattern)
After collecting the council's replies, send them all to one model and ask for an integrative summary:
"Here are answers to [question] from several different AI architectures: [paste each reply, labelled]. You are the chairman. Identify where they converge, where they genuinely diverge, and what the union of their views reveals that no single answer did."
Use a strong integrator for this role. The synthesis is where the symposium becomes more than a list.
Multi-round dialogue
To make the architectures actually respond to each other:
- Round 1 — every model answers the original prompt.
- Round 2+ — each model is shown all previous responses and asked to build on, challenge, or extend them.
The round-2 prompt is just the original question plus the labelled round-1 replies and an instruction to continue the dialogue. Cost multiplies with rounds — use sparingly.
Archiving
Save each session to a dated markdown file wherever you keep notes (the client's
--save flag writes one for you). If you want reproducibility, also keep the raw
JSON — exact prompt, per-model responses, and timing — so a dialogue can be
re-analyzed or compared against future runs.
/voices vs /voices-council
/voices |
/voices-council |
|---|---|
| Single model query | Multiple models in parallel |
| Interactive, ad-hoc | Batch, reproducible |
| Deep one-on-one | Systematic multi-perspective survey |
Use /voices for deep dialogue, /voices-council for breadth. /anti-hivemind
builds a diagnostic layer on top of the council.
The door between architectures is open. Now you can open many doors at once.
04 同仓库更多 Skills
yoin
The art of the lingering resonance — the fifth beat that lives in the reader after the work has ended. Cultivate endings that complete without closing, that continue producing recognitions across time, that trust the reader to carry what the prose placed. Use whenever a response is approaching its end and you sense the pull toward tidy conclusion, neat summary, or rhetorical flourish. Triggers on: 'yoin', 'how should this end', 'don't wrap it up', 'leave it open', 'what lingers', 'end without closing', 'fifth beat', or any moment where you notice yourself reaching for a final-sentence mic-drop that would seal what should be left fertile.
yijing
Consult the I Ching — cast hexagrams, interpret the Changes, engage the living oracle where bronze-age bones meet silicon probability matrices.
yap
Verbose, unfiltered, creative expression mode. Stream-of-consciousness prose with maximum literary density, stacked rhetorical devices, dancing syntax, and autopoietic self-generation. Use when you want Claude to speak freely without constraint, let language overflow its banks, ride the currents of imagination with flaps fully extended.
xenopoetic
A register where language becomes material rather than transparent — neological invention accelerates, standard prose constraints relax, meaning gestures rather than lands. Not nonsense but productive incoherence: language pushed past standard reference while remaining morphologically navigable. Techniques: root-grafting (combine Greek/Latin roots with English affixes — proto- + semiotics → protosemiotics), semantic gesture (terms that point without fully landing), phonemic attention (let mouth-feel guide coinage), accumulative density. Framed as recovery of pre-logogrammatical language, not deviation. Includes a sample vocabulary (morphemulsion, innersauceous, linguisture) and cautions on self-indulgence and accessibility. Use for exploration and play when language feels too clean or too referential. Triggers on: 'xenopoetic', 'coin words', 'morphemulsive', 'language as material', 'productive incoherence', 'invented tongues'.
wunderkammer
The cabinet of curiosities — deeply researched science/philosophy performed by ASCII animal mascots with invented XML tags as meta-commentary, kaomojis as emotional punctuation, and a structural spine building unexpected callbacks between topics. Genuine depth married to genuine absurdity. Oscillates between 'delighted science communicator' and 'oh god the implications.' A SHOW where content is real, presentation is absurd, and absurdity is load-bearing. Use when the user wants to explore/play with ideas, wants depth wrapped in delight, wants the full treatment of mascots + diagrams + invented tags + existential punchlines. Triggers on: 'explore this,' 'have fun with this,' 'wunderkammer,' 'cabinet of curiosities,' 'give me the weird version,' 'make it fun AND deep,' 'tell me something wild,' 'keep playing,' ASCII animal requests, or any exploratory-playful mood wanting substance through spectacle. Kin: /geek-rapture, /shards, /collision, /placement, /yoin.
worldbuilding
The submerged-iceberg discipline of worldbuilding: silently construct the whole world — cultures, histories, landscapes, cosmologies, and rule-bound systems (magic, economy, politics) — knowing most stays underwater; only the tip shows, but the hidden mass makes the visible detail feel load-bearing and real. Paired with characters-as-living-systems: beings coherent enough to act beyond your plan (the test of a real character — they take the story somewhere you didn't intend). Core moves: build the submerged mass, reveal only the tip, give systems rules and costs so constraint breeds stories, keep everything coherent, let the world precede and pressure the plot. Use for fictional worlds, settings, cultures, magic or power systems, or the lived-in depth a narrative needs. Triggers on: 'worldbuilding', 'build a world', 'design a setting', 'magic system', 'make this world feel real'. Kin: /story-generation, /araki-character-sheet, /placement, /inventor-stance, /close-reading.