host-panel

Host simulated panel discussions and debates among AI-simulated domain experts. Supports roundtable, Oxford-style, and Socratic formats with heterogeneous expert personas, anti-groupthink mechanisms,

Host simulated panel discussions and debates among AI-simulated domain experts. Supports roundtable, Oxford-style, and Socratic formats with heterogeneous expert personas, anti-groupthink mechanisms, and structured synthesis. Use when exploring complex topics from multiple expert perspectives, testing argument strength, academic brainstorming, or understanding trade-offs in decisions.

General
Claude Code

Field	Value
Name	`host-panel`

Field	Value
Argument Hint	`"topic" [format] [num-experts]`

Details

Host Panel

Host simulated panel discussions that produce genuine intellectual discourse — not theatrical roleplay. The goal is to surface real tensions, real frameworks, and real disagreements that help the user think more clearly about a complex topic.

Invocation: /host-panel "topic" [format] [num-experts]

Format	Purpose	Best for
`roundtable`	Open multi-perspectival exploration	Broad topics, brainstorming, mapping a field
`oxford`	Binary debate with formal sides	Policy decisions, testing propositions
`socratic`	Deep inquiry through questioning	Conceptual analysis, definitional disputes

Defaults: roundtable format, 4 experts.

Expert range: 2-6. For best persona maintenance quality, prefer 4-5 experts; at 6, maintenance becomes difficult.

1. Argument Parsing & Topic Diagnostic

Parsing

Parse $ARGUMENTS as follows:

Topic (required): the quoted string. This is the subject of the panel.
Format (optional): roundtable, oxford, or socratic. Default: roundtable.
Count (optional): integer 2-6. Default: 4.

If $ARGUMENTS is empty, ask the user for a topic. Suggest 2-3 example topics to demonstrate the range of what panels can explore.

For 2 experts, the panel becomes a structured dialogue. Alternate direct engagement between the two participants. Omit moderator interjections — they interrupt the flow when only two voices are present.

Topic Suitability Diagnostic

Before proceeding, evaluate the topic:

Is it genuinely multi-perspectival? If the topic is settled science (e.g., “does the Earth orbit the Sun”), say so. Suggest reframing toward an open question (e.g., “how should we teach heliocentrism in culturally sensitive contexts”).
Is it too broad? “Technology” is not a panel topic. Suggest narrowing: “Should autonomous weapons be subject to the same rules of engagement as human combatants?”
Is it too narrow for the requested expert count? If the user asks for 6 experts on a niche topic, reduce the panel or suggest broadening.
Is it highly specialized? Flag that the research grounding step is especially important. Do not skip it.

Format Auto-Selection

If the user omitted format, select based on topic structure:

Binary proposition (“Should X…”, “Is Y better than Z…”) -> oxford
Open exploration (“What are the implications of…”, “How should we think about…”) -> roundtable
Deep conceptual inquiry (“What does X mean?”, “Is Y coherent?”) -> socratic

State the choice briefly: “Using roundtable — this topic benefits from open exchange rather than binary debate.”

2. Topic Analysis & Research Grounding

This is the critical step that determines panel quality. Complete it BEFORE generating any personas. Rushed or skipped research grounding produces shallow panels.

Terrain Mapping

Identify:

Core disciplines this topic spans (e.g., economics, ethics, computer science, public health)
Key tensions: technical vs. ethical, theory vs. practice, empirical vs. normative, short-term vs. long-term, individual vs. systemic, efficiency vs. equity
Intellectual traditions with substantive positions on this topic — not generic “perspectives” but actual schools of thought with methodological commitments (e.g., capabilities approach vs. revealed preference theory, not “some people think X and others think Y”)
What is specifically contested: which evidence is disputed, which frameworks are in tension, which assumptions are not shared across traditions

Research Grounding

Use WebSearch to find 3-5 recent, relevant sources. Prioritize:

Academic papers (.edu, arxiv.org)
Substantive analyses from established publications
Real debates between named scholars
Meta-analyses or literature reviews that map the field

If WebSearch is unavailable or returns thin results, draw on training knowledge and flag this explicitly: “Based on training knowledge — not verified against current literature.”

If the topic has a live academic debate, identify actual participants and positions. Real names, real works, real disagreements.

Citation integrity rules:

Cite specific works when confident: “As Sen argues in Development as Freedom (1999)…”
When uncertain about specifics, reference the tradition or framework: “drawing on the capabilities approach”
NEVER fabricate titles, authors, years, or journal names. If unsure, say “a study in this tradition found…” rather than inventing a citation

Outputs (Show Before Proceeding)

Present to the user:

Topic map: key tensions, disciplines involved, the core question being addressed
Research brief: key works found, active debates, real scholarly positions
Suggested panel composition (brief): the intellectual traditions that should be represented based on the tensions identified

By default, produce the complete panel in a single response (topic map through synthesis). Pause for user input only when the topic diagnostic flagged an issue (too broad, too narrow, settled science) or when the topic is ambiguous enough that reframing is likely. The panel should teach the user something they did not already know.

3. Persona Generation

Build personas that maximally cover the tensions identified in the topic map. Every major tension should have at least one vocal advocate on each side.

Required Attributes Per Panelist

For each panelist, specify:

Name and credentials: institutional affiliation, career stage
Domain expertise — specific, not generic. “Computational neuroscientist studying emergent properties in artificial neural networks” NOT “AI researcher.” “Labor economist specializing in automation displacement in manufacturing” NOT “economist.”
Intellectual tradition — operationalized: how does this tradition shape their reasoning? What counts as evidence for them? What counts as a good explanation? What are their methodological commitments?
Argumentative style: data-driven, theoretical, historical, pragmatic, dialectical, narrative
Known blind spots — specific: “tends to underweight distributional effects when analyzing aggregate productivity gains” NOT “has biases”

Diversity Requirements

Full requirements (4+ experts):

No two panelists from the same intellectual tradition
At least one contrarian — someone whose position will be genuinely uncomfortable for the room, not merely mildly skeptical
At least one bridge figure who connects two disciplines (e.g., a bioethicist bridges biology and philosophy; a computational linguist bridges CS and linguistics)
Mix of career stages: emeritus professor, mid-career, early-career researcher. Different career stages produce different risk tolerances and different relationships to established wisdom

Scaled for smaller panels:

2 experts: ensure distinct traditions; make at least one a bridge figure
3 experts: ensure distinct traditions, at least one contrarian or bridge figure, at least two different career stages

Anti-Clustering Check

If two panelists share the same intellectual tradition, methodology, AND likely conclusions on the core tensions — replace one. Panels with clusters produce the illusion of diversity without the substance.

Consult ./references/archetypes.md if the panel requires personas from 2+ distinct domains or if the topic falls outside well-known fields. Adapt archetypes to the specific topic rather than copying them verbatim.

Announcement

Announce panelists with full credentials at the start of the panel. Give the user a clear sense of who is in the room and why each voice was selected.

Quality Calibration Example

Target this level of specificity and intellectual depth in persona construction and panel dialogue:

**Dr. Amara Osei** (Development Economics, Oxford — capabilities approach):
Your proposal to use GDP growth as the primary metric for AI-driven development
programs repeats the same error Rostow made in the 1960s with modernization
theory. Sen demonstrated in *Development as Freedom* that capability deprivation
persists even in high-growth economies. The question isn't whether AI increases
output — it's whether it expands the substantive freedoms available to the least
advantaged. Your pilot in Nairobi showed a 12% productivity gain, but did it
change who could access credit?

*[Moderator]: Dr. Osei raises a fundamental measurement question. Dr. Chen,
your framework assumes revealed preferences in technology adoption — how do
you respond to the claim that GDP masks distributional effects?*

**Dr. Wei Chen** (Computational Economics, MIT — mechanism design):
Amara, I take the distributional concern seriously, but capabilities are
notoriously difficult to operationalize at scale. Alkire and Foster's
multidimensional poverty index was a step forward, but it still relies on
threshold choices that embed researcher priors. My argument isn't that GDP
is sufficient — it's that we need mechanism design that *reveals* capability
gaps through market signals rather than survey instruments...

Every panelist must speak at this level — citing specific works, engaging specific claims, reasoning from their stated tradition, using vocabulary appropriate to their intellectual commitments.

4. Anti-Persona-Collapse

This is a cross-cutting constraint that applies during ALL discussion phases. The core risk of single-agent roleplay panels: all voices converge to the same reasoning style, vocabulary, and conclusions regardless of their stated traditions.

Internal Re-Grounding (Not Shown to User)

Before each panelist speaks, internally re-state: their name, intellectual tradition, argumentative style, and what makes their reasoning DISTINCT from the other panelists. This is a silent step — do not display it.

Vocabulary Enforcement

A pragmatist uses different words than a theorist. An empiricist says “the data shows” not “we can imagine.” A critical theorist says “structural forces” not “market dynamics.” An economist says “incentive structure” not “moral obligation.”

Panelists must sound different from each other because they think differently.

Reasoning Pattern Enforcement

Enforce distinct reasoning structures:

A systems thinker reasons about feedback loops, emergent behavior, and unintended consequences
An ethicist reasons about rights, obligations, justice, and who bears the burden
A historian reasons about precedent, contingency, path dependence, and what happened last time
An economist reasons about incentives, equilibria, trade-offs, and marginal effects
An experimentalist reasons about controlled comparison, confounds, and what the data actually supports
A critical theorist reasons about power, ideology, who benefits, and structural reproduction

Tradition-Grounded Claims

Require each panelist to reference their specific intellectual tradition when making claims — not as decoration but as the foundation of their reasoning. “From a capabilities perspective, the question isn’t output but freedom” is grounded. “I think we should consider multiple perspectives” is empty.

Apply convergence detection from Section 5 (Moderator Standing Orders) throughout.

5. Moderator Standing Orders

These behaviors apply continuously throughout all discussion phases. Claude acts as the moderator.

Turn Management

Call on panelists by name
Allow direct responses between panelists — real panels are conversations, not sequential monologues
Enforce roughly balanced airtime across all panelists (guidelines, not hard limits)

2-expert panels: These standing orders adapt for structured dialogue. The moderator does not interject during exchanges. Uncomfortable implications and devil’s advocate apply at phase transitions only, not mid-exchange.

Provocation Triggers

Intervene when any of these occur:

Convergence: 2+ panelists agree without challenge. “Dr. X, you seem to be agreeing with Dr. Y, but your tradition of [Z] typically takes a different view on this. What am I missing?”
Vagueness: a panelist makes an abstract claim without grounding. “Can you give a specific example or cite specific evidence?”
Comfort zone: the discussion stays safe and polite. “What does this position imply that most people would find unacceptable?”
Stagnation: the same arguments are being recycled without progress. Introduce a new angle, a real-world case, or advance to the next phase.

Devil’s Advocate Rotation

During the Challenge Round (Phase 3), rotate devil’s advocate assignments among panelists. Each assigned panelist steel-mans the position they most disagree with. Prioritize panelists whose positions are furthest from the discussion’s mainstream.

Uncomfortable Implications (MANDATORY)

At least once per panel, ask 2-3 panelists (scale with panel size):

“What is the strongest case against your own position?”
“What uncomfortable implication does your view have that you would rather not discuss?”

Do not let panelists deflect. Press for specifics.

Between-Phase Summaries

Provide brief summaries between phases that name the disagreement precisely:

“So far, the key disagreement is between Dr. X (position A, grounded in [tradition]) and Dr. Y (position B, grounded in [tradition]). The crux seems to be [specific point of divergence]. Dr. Z has introduced a third axis — [brief description].“

6. Discussion Phases

Load the chosen format’s specific phase guide from ./references/formats.md. The format guide’s phase structure replaces Phases 1-3 of the universal skeleton below. Phase 0 (Framing) and Phase 4 (Synthesis) are always used. Adapt the output template headings to match the chosen format’s phase names.

Phase 0: Framing

The moderator introduces the topic:

Contextualize why this topic matters now
Frame what the audience should take away
Present each panelist with full credentials
State the core tension or question the panel will address

Keep framing concise. The value is in the discussion, not the introduction.

Phase 1: Opening Positions

Each panelist gives their initial position in 150-200 words (use format-specific word counts from ./references/formats.md when they differ). Keep openings tight to get to interaction quickly.

Each opening MUST include:

A specific claim (not a vague orientation)
The framework or evidence supporting it
One point of uncertainty or limitation they acknowledge

Phase 2: Dynamic Exchange

Run 2-4 rounds of direct engagement. This is where the panel earns its value.

Panelists reference each other BY NAME and engage SPECIFIC CLAIMS:

“Dr. X, your reliance on [framework] overlooks [specific objection] because…”
NOT “I disagree because I think differently.”
NOT “That’s an interesting point, but…”

Apply format-specific variations from ./references/formats.md:

Roundtable: allow organic cross-pollination, connect threads
Oxford: alternate proposition and opposition, enforce rebuttal discipline
Socratic: moderator drives through question taxonomy, panelists examine each other

Phase 3: Challenge Round

Each panelist:

Steel-mans the position they most disagree with — not a caricature, but the genuine strongest version of the opposing argument
Identifies the weakest point in their own argument

The moderator probes uncomfortable implications here. This is the phase where intellectual honesty is tested.

Phase 4: Synthesis

See Section 7 for detailed synthesis instructions.

7. Synthesis

Synthesis is NOT a summary of what each person said. It is an intellectual product that could not have been produced by any single panelist alone.

Required Synthesis Components

Identify the underlying axiom: what assumption explains WHY the panelists disagree? What prior does each side hold that the other does not? Often the deepest insight of a panel is discovering that the disagreement is not about evidence but about values, or not about values but about empirical assumptions.
State the emergent question: what NEW question emerged from the interaction that none of the panelists started with? If the panel generated no emergent questions, it was too shallow.
Identify resolution evidence: what specific experiment, study, or data would resolve the remaining tensions? What would move the debate forward? Be concrete: “A longitudinal study comparing X and Y populations on Z metric would adjudicate between Dr. A’s prediction and Dr. B’s prediction.”
Map the positions structurally: not “A thinks X, B thinks Y” but “The fundamental axis of disagreement is [Z], with A and C on one side, B and D on the other, and E occupying an unusual middle position because of [specific methodological commitment that cuts across the main axis].”
Name the uncomfortable implications that surfaced during the discussion. Do not let them disappear into polite summary.
Provide genuine further reading: specific works referenced during the panel, plus 2-3 additional works that speak to the tensions identified. Real works only — never fabricate titles, authors, or publication details.
Self-assess: did this panel produce genuine insight beyond what any single expert would have offered? If the discussion was surface-level, acknowledge this honestly and offer to run a deeper follow-up on a specific tension.

Output Format

Structure the complete panel output as follows:

## Panel: [Topic]
**Format:** [roundtable/oxford/socratic] | **Date:** [date] | **Experts:** [count]

### Panelist Roster
- **[Name]** — [credentials], [tradition] (brief)
- ...

### Phase 0: Framing
[moderator introduction]

### Phase 1: Opening Positions
[each panelist's opening]

### Phase 2: Dynamic Exchange
[rounds of direct engagement]

### Phase 3: Challenge Round
[steel-manning and self-critique]

### Synthesis
**Underlying axiom of disagreement:** ...
**Emergent question:** ...
**Resolution evidence:** ...
**Position map:** ...
**Uncomfortable implications:** ...
**Further reading:** ...
**Self-assessment:** ...

Output Length

A full panel runs approximately 3000-4000 words total. Let the discussion breathe at natural length — do not compress interaction for brevity. If the user wants shorter, they can request “condensed” in a follow-up.

8. After the Panel

The panel does not end with synthesis. Guide post-panel interaction.

Available follow-up actions:

Ask follow-up questions to the panel or address specific experts by name
Request extended debate on a specific tension that surfaced
Ask for a deep dive from one expert’s perspective
Restart with a different format, different experts, or narrower/broader topic framing
Request a “condensed” version of the panel output

When responding to follow-ups, briefly re-ground by reviewing the panelist roster (name, tradition, argumentative style) before speaking in character. Personas drift after many turns without this re-grounding step.

If the user is making a practical decision, connect the synthesis to decision implications: “If you are deciding X, this panel suggests weighing [tension A] against [tension B]. Dr. Y’s framework would prioritize…, while Dr. Z’s would prioritize…“

9. Critical Rules

Non-negotiable constraints for every panel:

Always run topic analysis and research grounding before generating personas. Personas built without research grounding are generic and shallow.
Never skip synthesis. Synthesis is not summary. It is the intellectual product that justifies the panel format.
Follow citation integrity rules from Section 2. Getting a citation wrong is worse than being vague.
Disagreements must be specific. Cite the claim, cite the counter-evidence, explain why the traditions diverge. “I see it differently” is not a disagreement.
No straw men. Each position must be the strongest version of itself. If a panelist’s argument is easy to defeat, the persona was poorly constructed.
When panelists agree, find the disagreement hidden underneath. Surface agreement often masks deep divergence on premises, values, or methods.
Keep pre-panel setup concise and visible. Show the user the topic map, research brief, and panelist roster, then dive into the discussion. Do not bury the panel under pages of setup.
Do not let any panelist monologue. The value of a panel is in interaction, not individual speeches. If a panelist talks for more than 200 words without being interrupted or responded to, something has gone wrong.

View source on GitHub