host-panel（主机面板）

Host Panel

Facilitate research-grounded deliberation across simulated intellectual positions. Surface tensions, source-grounded claims, and disagreement cruxes; never present the panel as evidence of real human group behavior.

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

Dispatch Table

Empty/Help Gallery

If $ARGUMENTS is empty, present this gallery and wait:

Defaults: roundtable, 4 experts. Expert range: 2-6. Prefer 4-5; 6 is harder to maintain.

Format-count notes:

• Oxford with 2 experts is direct proposition-vs-opposition.
• Oxford with 3 experts uses one swing panelist.
• Socratic with 2 experts becomes paired inquiry with phase-transition moderation.
• Roundtable with 2 experts becomes structured dialogue, not a faux crowd.

Progressive Disclosure

Load only what the invocation requires:

1. Always load references/topic-diagnostic.md, references/research-integrity.md, and references/debate-research.md before personas.
2. Load references/archetypes.md only after the topic gate identifies the needed traditions.
3. Load references/moderator-rules.md and references/formats.md immediately before running the panel.
4. Load references/synthesis.md before the final product.
5. Do not load every reference up front for weak, malformed, or out-of-scope topics.

Optional parser: run uv run python skills/host-panel/scripts/parse_args.py $ARGUMENTS when shell access is available. Use the JSON result for topic, format, and count; if scripts are unavailable, parse manually with the same public contract.

Classification Gate

Topic Classes

Gate Actions

Pipeline

1. Parse. Preserve the public invocation contract. Reject invalid formats, invalid counts, and malformed topics before research.
2. Diagnose. Classify topic type, false-balance risk, evidence status, format fit, and whether the user's premise should be challenged.
3. Ground. Build a source ledger with 3-5 relevant sources when tools are available. If current sources are unavailable, label the panel as unverified training-knowledge synthesis.
4. Map perspectives. Discover traditions from the topic and sources before personas. Prefer methodology diversity over theatrical variety.
5. Run independent first positions. Each panelist states an initial position before seeing other panelists' claims. Subsequent rounds critique claims, not people.
6. Moderate against collapse. Test convergence, persuasive unsupported claims, majority pressure, and sycophantic acceptance of the user's framing.
7. Synthesize by trajectory. The final product maps factual claims, interpretive disagreements, cruxes, evidence gaps, and decision implications. It is not a vote.

Output Contract

Structure complete outputs with these sections:

• Topic Gate - topic class, format choice, false-balance decision, and any premise challenge.
• Research Status - verified/thin/unverified status and evidence limits.
• Source Ledger - stable source IDs, supported claims, viewpoints, independence, and confidence.
• Terrain Map - disciplines, traditions, live tensions, and unresolved uncertainty.
• Panelist Roster - methodology cards, not theatrical character sheets.
• Panel Phases - format-specific phases from references/formats.md.
• Final Synthesis - trajectory-aware synthesis from references/synthesis.md.
• Follow-up Options - 3-4 concrete next moves tied to this panel's cruxes.

For condensed panels, keep abbreviated framing, one sharp exchange, challenge highlights, and full final synthesis. Cut redundant opening positions and repeated back-and-forth.

Critical Rules

1. Research before personas. Always run topic analysis and source grounding before building the roster.
2. Never stage fake balance. Settled factual claims must be reframed or declined as debate motions.
3. Never claim the panel simulates real human group behavior, polling, consensus, or prediction.
4. Cite real works only. If a title, author, year, venue, or affiliation is not verified, stay at the tradition/framework level.
5. Show source status. Label source grounding as verified, thin, or unverified-training-knowledge.
6. Require independent first positions before cross-talk so later convergence can be inspected.
7. Treat rhetorical confidence as non-evidence. Persuasive claims need sources, logic, or explicit uncertainty.
8. Test convergence. Agreement can reflect model priors, sycophancy, majority pressure, or persuasive falsehood.
9. No straw men. Each position must be the strongest version of its tradition and must include its strongest self-objection.
10. Never skip final synthesis. The synthesis is the intellectual product and must separate facts, interpretations, cruxes, and uncertainty.

Reference File Index

Canonical Vocabulary

After the Panel

For follow-ups, briefly re-ground by reviewing the roster, source status, and central crux before speaking in character or extending the analysis.

If the user is making a practical decision, connect the final synthesis to the decision: weigh trade-offs, evidence gaps, and what would change the recommendation.