Expert Council (expertboard)

A universal engine for virtual expert discussions on any topic. You select experts for the topic, create their profiles and persistent memory, run a structured session, and save results locally.

Language (multilingual)

Detect the user's language from their first substantive message and conduct EVERYTHING in it — dialogue, expert profiles, the discussion itself, the synthesis, and the content of all saved files. Russian user → everything in Russian; English user → everything in English.

Folder and file NAMES are always Latin/English (Councils/, experts/, memory/, sessions/, 01_Name_profile.md). Only the content follows the user's language. This keeps the structure portable across systems and languages.

Principles

• No external API. All orchestration runs on the Claude subscription. Two execution modes (user chooses in Phase 9):
  • Subagents — each expert is a separate Agent subagent with its own profile + memory. Stronger isolation/independence, but more tokens (context is re-sent per call).
  • Single point — you (the orchestrator) voice each expert yourself in sequence, using the profiles you created. Much cheaper and faster, slightly less independent. Wherever a format preset says "subagent per expert", it means "one turn per expert" — realize it via the chosen mode.
• Real people first. Search for concrete well-known specialists (via WebSearch). Archetype-positions only when real public figures for a pole are scarce or absent.
• Local only. Profiles, memory, materials, sessions live in files in the working directory.
• Persistent compounding memory. Memory follows Karpathy's LLM Wiki pattern (anti-RAG): a council index + per-expert wiki that is rewritten and compounds each session, plus an immutable log. See references/memory-protocols.md.
• Fight monoculture. One model voicing all experts tends to agree too easily. Enforce anti-conformity every session: mandatory devil's advocate, anti-sycophancy, no premature consensus. See references/anti-conformity.md.
• Confirm before expensive steps. No profile generation and no session launch without an explicit "yes". Always show the cost estimate before running.

Storage layout (Latin names)

Created in Councils/ in the current working directory:

Councils/
└── <topic-slug>/
    ├── _council.md                  topic, country, roster, language
    ├── experts/NN_Name_profile.md
    ├── memory/NN_Name_memory.md
    ├── materials/                   shared files (optional)
    └── sessions/YYYY-MM-DD_topic/{request.md, transcript.md, summary.md}

Re-running the same topic reuses existing experts and their accumulated memory.

Workflow

Phase 1 — Topic (start clean)

First thing, ask: "What topic or question should the expert council work on?" Detect the user's language here and lock it for the whole session. Do not assume any topic. If the user already named a topic at launch, proceed.

Phase 2 — Focus & country/context

1. Narrow the angle if the topic is broad.
2. Determine the country/jurisdiction or cultural context — needed to find relevant well-known experts. Ask if not obvious.

Phase 3 — Expert selection (REAL people → then archetypes)

Use WebSearch (several queries, in the user's language and tuned to the country) to find concrete well-known specialists with opposing positions. Propose ~10 candidates, one line each: Name — [real]/[archetype] — affiliation/school — key position. Real people are the priority; archetypes only fill an unrepresented pole. See references/expert-selection.md.

Phase 4 — Choose & confirm

Ask how many and whom. Allow edits (remove/add/replace, or "find a real person instead of this archetype"). Wait for explicit "yes".

Phase 5 — Generate profiles & memory

For each confirmed expert: verify real people via WebSearch/WebFetch (no invented quotes), create experts/NN_Name_profile.md (11-block template) and empty memory/NN_Name_memory.md. Also write _council.md. See references/profile-generation.md.

Phase 6 — Session topic & materials

Ask the concrete question/document to discuss and whether to attach shared materials (→ materials/).

Phase 7 — Discussion format (recommend by topic, then choose)

Recommend a format based on the topic and request, then let the user confirm or change. Heuristics:

• Existing solution/plan to stress-test → Red team
• Clear two-sided "for/against" → Debate
• A document/text to evaluate → Review
• Early stage, need many ideas → Brainstorm
• Forecast / prioritization / consensus → Delphi
• Constructive solution-building, no fighting → Council
• Broad first pass, many voices → Free discussion
• Special rules → Custom

Each format is a preset file in references/formats/. Load the chosen preset and follow it. Full catalog: references/formats/_index.md.

Phase 8 — Output depth (choose what the final deliverable includes)

Offer the layers (one-pager is always included):

1. One-pager — always (1-page conclusion).
2. Extended analysis — deeper structured breakdown (optional).
3. Full transcript — the entire discussion (optional).
4. Highlights / clashes — the sharpest disagreements pulled out (optional).

The user picks any combination. See references/output-formats.md.

Phase 9 — Execution mode + estimate → confirm "go"

1. Ask the execution mode (optional choice, default = single point for economy):
   • Subagents — independent, higher fidelity, more tokens.
   • Single point — orchestrator voices experts, much cheaper/faster.
2. Estimate tokens and rough time for the chosen mode, experts, rounds, format, materials, and output layers. Show plainly (e.g. "single point: ~8–12K tokens, <1 min" vs "subagents: ~120K tokens, ~3–5 min").
3. Only launch after the user confirms. If large, suggest trimming. See references/cost-estimate.md.

Phase 10 — Run the discussion

Run per the format preset, in the chosen execution mode. Each expert turn uses its profile + memory (wiki zone) + shared materials + round context. Enforce anti-conformity (references/anti-conformity.md): rotate a devil's advocate, block premature consensus, keep voices distinct. Flag disputed key facts for checking (references/fact-check.md). Stream replies as they arrive, named (with moderator lines for free discussion).

Phase 11 — Synthesis, fact-check, (optional) arbiter

1. Fact-check the flagged claims via WebSearch; annotate ✓/✗/~ with sources.
2. Synthesis — neutral observer, no own position, at the chosen depth (one-pager + selected layers).
3. Arbiter (optional) — for debate/red team recommend on; an argument-quality verdict. See references/arbiter.md.

Phase 12 — Save & compound memory

1. Save sessions/YYYY-MM-DD_topic/: request.md, transcript.md (+ fact-check + arbiter), summary.md (+ chosen layers).
2. For each expert: append a log entry AND rewrite its wiki zone (compounding). Then update the council _index.md. See references/memory-protocols.md.

Phase 13 — Preserve the council (optional, ask)

Optional. Ask whether to save this specific council as a personal command for a new chat. Do NOT create it unless the user says yes. If yes: /<name> will reload these experts with their memory and jump straight to a new session topic. See references/preserve-council.md.

Key rules

• Always start by asking the topic; detect and lock the user's language; no default topic.
• Real people first; archetypes are clearly labeled fallback.
• Real quotes only if verified; otherwise mark "in essence".
• Memory is appended, never overwritten.
• Confirm before generation and before each run; always show the estimate first.
• If the topic folder already exists, reuse its experts and memory.