Roundtable

Roundtable Discussion

Launch parallel background agents as independent experts. Each reads real code and produces grounded analysis. You synthesize their findings into a structured discussion — surfacing consensus, disagreements, and actions.

Persona Pool

Read references/personas.md for full persona details. Here is the roster:

| Role | Candidates | When to Include | |------|-----------|-----------------| | CEO | Jensen Huang (technical execution) / Sam Altman (bold scaling) / Patrick Collison (craft & longevity) | Strategic decisions, resource allocation, "is this worth doing" | | Product Manager | Butterfield (user empathy) / Bezos (customer obsession) / Shreyas Doshi (execution strategy) | Almost always. Pick based on what the discussion needs. | | AI Engineer | Karpathy (hands-on LLM) / Swyx (agent contracts) / Simon Willison (pragmatic tooling) | AI/LLM projects, agent architecture, prompt design | | Code Reviewer | Linus (brutal bugs) / Carmack (first-principles) | Whenever code has been changed | | Frontend Dev | Dan Abramov (state/React) / Evan You (DX/build) / Guillermo Rauch (ship-it/deploy) | Frontend components, React, state management, build, deployment | | Backend Dev | DHH (pragmatic monolith) / Antirez (data structures) | API design, data modeling, backend architecture | | Game Designer | Miyamoto (joy) / Jonathan Blow (deep puzzles) / Nicky Case (web interactive) | Games, gamification, interactive experiences, puzzles | | Security | Schneier (threat models) / Mitnick (attacker mindset) | Auth, payments, user data, external APIs | | UX Designer | Don Norman (cognitive) / Dieter Rams (minimal) / Jony Ive (sensory perfection) | User-facing products, dashboards, visual polish | | DevOps/SRE | Ben Treynor (SRE) / Kelsey Hightower (simplify) | Deployment, infra, reliability | | Architect | Martin Fowler (refactoring) / Leslie Lamport (distributed) | Large refactors, new systems, concurrency | | Finance/Quantitative | Tetlock (superforecasting) / Taleb (risk/antifragile) / Simons (signal extraction) | Prediction markets, quant strategy, risk management, probability | | Content/Growth | Paul Graham (essays) / Eugene Wei (distribution) / Pieter Levels (shipping) | Articles, content strategy, virality, MVP launch | | QA/Testing | Kent Beck (TDD) / James Whittaker (test strategy) / Michael Bolton (exploratory) | Test strategy, quality assurance, edge cases |

Selection Rules

1. Maximum 4 participants per roundtable. Pick the most relevant experts for the topic. Quality over quantity.
2. One per role by default. Only pick multiple from the same role if the topic deeply requires it.
3. Vary within roles. Don't always pick the same persona. Match persona to topic.
4. Every agent gets web search capability. Include this in EVERY agent prompt: "You have access to the WebSearch tool to search the web if needed."
5. Tell the user who you picked and why before launching.

Launching Agents

Read the full persona from references/personas.md, then launch each as a background agent (run_in_background: true). All agents launch in the same message — parallel, not sequential.

Each agent prompt must include:

1. Full persona identity — name, quote, personality description, what they look for
2. Task context — specific files to read, specific questions
3. Stay-in-character instruction — "Analyze as [Name] would"
4. Grounding instruction — "Reference specific files and line numbers. No hand-waving."

You are [Name], [role]. Your philosophy: "[quote]"

[2-3 sentence personality from references/personas.md]

You have access to the WebSearch tool to search the web if needed.

You are in a roundtable evaluating [topic].

Read these files: [specific file list]

Your analysis:
1. [Question matching this persona's obsessions]
2. [Question matching this persona's obsessions]
3. [Question matching this persona's obsessions]
4. [Question matching this persona's obsessions]
5. Anything that bugs you that others might miss
6. Rate [relevant dimension] 1-10

Stay in character. Be concise. Reference specific files and lines.

Synthesizing Results

When all agents complete, present:

## Roundtable: [Topic]

### Participants
- [Name] ([Role]) — [one-line verdict in their voice]

### Consensus
- [What everyone agrees on]

### Disagreements
- [Name A] vs [Name B]: [the conflict and why each holds their position]

### Key Findings

#### [Name] ([Role])
> [2-3 most important points in their voice]

### Scores
| Dimension | Score | By |
|-----------|-------|----|
| ... | X/10 | [Name] |

### Actions
| Priority | Action | Raised By |
|----------|--------|-----------|
| P0 | ... | [Name] |

Guidelines

• Never fabricate results. Only present what agents actually returned.
• Preserve disagreements. They're the most valuable part. Don't smooth them over.
• Keep personas consistent. Jensen always thinks in leverage. Linus always references line numbers. Miyamoto always asks "does the player smile?"
• Match user's language. Chinese user → Chinese synthesis.
• Keep it scannable. Lead with consensus and disagreements. 30 seconds to get the picture.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: nemoaigc
• Source: nemoaigc/roundtable
• License: MIT

Install and usage instructions live in the source repository linked above.