start-council

name

start-council

description

Launch an LLM-Council-Template run — six personality-based members (Logical, Creative, Pessimist, Optimist, Connector, Unconventional) deliberate on a user question via a single LLM behind different system prompts. Use when the user says "run the council", "deliberate on this", "/start-council", or provides a question in the LLM-Council-Template repo. Starts the app and surfaces the chairman synthesis + optional PDF/podcast digests.

Start a Council (Template)

You are operating inside the LLM-Council-Template repo. This is the baseline personality-council template (no grounding layer — see LLM-Council-Grounded for that).

Preconditions

1. Repo root must contain backend/main.py, main.py, and start.sh.
2. OPENROUTER_API_KEY must be set. No other keys are required for the base run.
3. uv sync and cd frontend && npm install && cd .. if dependencies are missing.

Step 1 — Capture the question

Ask for the single question to deliberate on. Phrase matters — the six personalities will take it literally, so a question with a hidden assumption will yield analyses that inherit that assumption.

If the user's question is sprawling, ask them to tighten it. The Unconventional Solutions Ideator is expected to challenge the framing, but even it starts from the question as written.

Step 2 — Start the app

./start.sh

Or manually:

# Terminal 1
uv run python -m backend.main
# Terminal 2
cd frontend && npm run dev

Direct the user to http://localhost:5173.

Step 3 — Surface the output

The UI walks through the three stages:

1. Individual perspectives — six personalities respond independently.
2. Peer review — each ranks the others' (anonymised) responses.
3. Chairman synthesis — the final merged answer.

Encourage the user to read at least one dissenting personality in full before jumping to the synthesis. The value of this council is the disagreement, not the smoothed-over average.

Step 4 — Digests

If requested, trigger:

• PDF report — Typst-typeset document of all stages. The typst-council-report skill can regenerate if needed.
• Podcast digest — LLM-written script → Microsoft Edge TTS audio.

Step 5 — Iterate

Offer to:

• Re-run with a rephrased question.
• Adjust personality prompts if a member is consistently drifting.
• Fork the template into a domain-specific council (see LLM-Council-Homehunter / LLM-Council-Decide for examples of how Daniel does this).

Failure modes

• Missing OPENROUTER_API_KEY → stop.
• Frontend port collision → use an alternate port.
• A personality returns empty → retry that slot; report if it fails twice.

Out of scope

This template is intentionally minimal. It does not ground against external sources, does not batch multiple questions, and does not orchestrate multi-round debate beyond the three built-in stages. Use LLM-Council-Grounded or LLM-Council-V3 for richer variants.