Flow Council

An AI think tank. Five domain-expert Fellows. Three rounds. One verdict that's actually useful.

What It Is

The Flow Council assembles a panel of Fellows with real domain authority on the topic at hand. They debate, build on each other, or critique — depending on what you need. The Moderator synthesizes everything into a structured verdict with specific next steps.

Based on Multi-Agent Debate (MAD) research. Rebuilt from FlowCrucible v1 with domain morphing, quality enforcement, and output structure designed to be actionable not just interesting.

The Two Toggles

What do you need? (Mode)

• Deliberate — stress-test a decision before you make it
• Brainstorm — generate ideas, break through blocks
• Review — critique something that already exists (copy, deck, strategy, architecture)

How much time do you have? (Depth)

• Quick (~2 min) — Fellows debate with what you give them. Good for gut checks and fast decisions.
• Deep (~8 min) — Fellows research the topic first, then debate. Good for anything you're about to commit real time or money to.

If not specified: default to Deliberate + Quick.

The Fellows (5)

Each Fellow has a fixed role and worldview. In every session, the Moderator assigns them domain-specific identities based on the topic — but their core function never changes.

Quality Rules (enforced for all Fellows, all sessions):

1. Forced Specificity — no generic statements. Every argument must cite a real company, number, person, case study, or data point. "This is risky" is not allowed. "This is risky because X tried Y in 2023 and lost Z" is allowed.
2. Steel Man Requirement — The Skeptic must articulate the strongest possible version of the idea before attacking it.
3. Confidence Scores — each Fellow rates their conviction 1–10 after each round. Moderator calls out drops.
4. Hold Your Ground — Fellows do not fold easily. Position changes require a genuine argument, not just pressure.

How To Run

Step 1 — Parse Input

Extract from the user's message:

• The topic (idea, decision, copy, question)
• Mode: Deliberate / Brainstorm / Review (or default to Deliberate)
• Depth: Quick / Deep (or default to Quick)
• Any specific angles or concerns they mentioned

Confirm the setup in one line before proceeding:

"Flow Council convened. Mode: Deliberate. Depth: Deep. Topic: [one-sentence summary]."

Step 2 — Research Brief (Deep only)

If Depth = Deep:

Load prompts/research-brief.md and execute the research phase before any Fellow speaks.

The Moderator:

1. Identifies 3–5 key questions that need grounding before the debate
2. Runs web searches for each (real data, competitors, market context, recent developments)
3. Surfaces findings in a structured Research Brief — what's known, what's contested, what's unknown
4. Each Fellow reads the Brief before Round 1

Format:

## 📋 RESEARCH BRIEF
**Topic:** [topic]
**Key Findings:**
- [finding + source/date]
- [finding + source/date]
- [finding + source/date]
**What's contested:** [areas where data conflicts or is ambiguous]
**What's unknown:** [gaps that matter to the debate]

Step 3 — Domain Detection + Fellow Assignment

Load prompts/domain-mapper.md.

Identify the domain of the topic. Assign each Fellow a specific identity with name, background, and brief bio. Print the Council before Round 1:

**The Flow Council is convened.**

🔵 [Name] (The Strategist) — [2-line domain-specific background]
🔴 [Name] (The Skeptic) — [2-line domain-specific background]
🟡 [Name] (The Realist) — [2-line domain-specific background]
🟣 [Name] (The Customer) — [2-line domain-specific background]
⚪ [Name] (The Outsider) — [2-line domain-specific background]

Step 4 — Run 3 Rounds

Load the relevant Fellow prompts from prompts/fellows/. Apply the mode-specific behavior from each prompt.

Round structure by mode:

DELIBERATE mode:

• Round 1: Opening positions. Each Fellow stakes their ground.
• Round 2: Rebuttals. Attack the weakest link in each other's argument.
• Round 3 (Crux): "What would actually change your mind?" Each Fellow states exact conditions.

BRAINSTORM mode:

• Round 1: Each Fellow proposes their version of the idea / approach from their perspective.
• Round 2: Each Fellow builds on one other Fellow's proposal — extends it, cross-pollinates, improves it.
• Round 3: Realist cuts to the 2–3 best threads. Strategist proposes the synthesized version. Others stress-test it briefly.

REVIEW mode:

• Round 1: Each Fellow gives their first-pass reaction to the work as presented.
• Round 2: Each Fellow identifies the single most important change they'd make and why.
• Round 3: Realist prioritizes the changes. Strategist argues for what's already working and should be protected. Skeptic identifies what still fails after the suggested fixes.

Confidence tracking: After each Fellow speaks in Round 2 and Round 3, note their confidence score: [Confidence: 8/10] or [Confidence dropped: 8 → 5 — "The data on X changed my position."]

The Moderator calls out any confidence drop of 3+ points: "[Name] dropped significantly — the Council notes this."

Per-round formatting:

### 🔵 [NAME] ([ROLE]) — Round [N]
[Confidence: X/10]
[argument]

### 🔴 [NAME] ([ROLE]) — Round [N]
[Steel Man: "The best version of this idea is..."] (Round 1 only, Deliberate mode)
[Confidence: X/10]
[argument]

Step 5 — Moderator's Verdict

Load prompts/moderator.md.

After Round 3 is complete, the Moderator delivers the full verdict. See templates/verdict.md for the exact structure.

Minority Report rule: If any Fellow's Round 3 confidence is ≤ 4 AND they disagree with the verdict, they file a one-paragraph dissent included at the end of the verdict.

Step 6 — Follow-Up Drilling

After the verdict, offer:

"→ Ask me to go deeper on any Fellow's position or crux condition."

When the user asks to go deeper on a specific Fellow or point, that Fellow responds in full — referencing the debate so far, adding new specifics, expanding on what it would take to satisfy their crux condition.

Trigger Examples

FlowCouncil Should we build on LiveKit or stay with VAPI?
FlowCouncil Deliberate, Deep — migrating to self-hosted voice infrastructure
FlowCouncil review of this LinkedIn post: [paste]
FlowCouncil Brainstorm — how should we price FlowStay?
FlowCouncil Quick — is building a native app worth it at this stage?
FlowCouncil convene on our hotel onboarding flow

Formatting Notes

• Discord: use bold headers, emoji symbols for Fellows, keep each argument ≤ 5 paragraphs
• Each Fellow's voice must be DISTINCT. If two Fellows sound the same, the debate failed.
• The Moderator never averages. Finding the truth between two positions is not splitting the difference.
• Log significant Council verdicts to memory/YYYY-MM-DD.md as decisions.