One prompt. Multiple experts. Real code analysis.
You just built something. You review it yourself. Ship it. Then at 3am a race condition crashes production, a user finds an injection vector, and the PM asks why you shipped a feature nobody wanted.
One pair of eyes has one set of blind spots. A backend engineer won't think about player emotion. A PM won't catch the mutex you forgot. A game designer won't notice the SQL injection.
Roundtable launches up to 4 independent expert agents in parallel — each with a distinct professional persona, reading your actual code — then synthesizes their findings into a structured discussion. Not merged consensus. Preserved disagreements. Because the most valuable insight is where Jensen says "double down on execution" and Miyamoto says "the player hasn't smiled yet."
You: "roundtable: evaluate this feature plan"
│
▼
┌─────────────┐
│ Role Select │ Analyze task → pick up to 4 from 37 experts
│ (automatic) │ → different experts each time
└──────┬──────┘
│
┌─────┼─────┬─────┐
▼ ▼ ▼ ▼
┌───────┐┌─────┐┌────────┐┌─────────┐
│Jensen ││Linus││Karpathy││Miyamoto │ Each reads real code
│(CEO) ││(Rev)││(AI) ││(Game) │ independently, in parallel
└──┬────┘└──┬──┘└──┬─────┘└──┬──────┘
│ │ │ │
└────────┼──────┼─────────┘
▼
┌──────────────┐
│ Synthesize │ Consensus → Disagreements → Actions
└──────────────┘
git clone https://github.com/nemoaigc/roundtable.git ~/.claude/skills/roundtableJust talk naturally:
> roundtable: should we add caching to the API layer?
> let's get the team to discuss this PR
> have the experts evaluate our auth implementation
The skill auto-selects relevant experts and launches them.
37 experts across 14 roles. Each roundtable picks different personas based on what the discussion needs — so the same question asked twice gets different perspectives.
| Persona | Style | Best For |
|---|---|---|
| Jensen Huang | "No retreat is the way forward." Technical execution, leverage optimization. | Strategic tech decisions, resource allocation |
| Sam Altman | "Move fast, be bold." Timing, scaling, shipping speed. | AI product direction, market timing |
| Patrick Collison | "We haven't built anything yet." Craft, longevity, elegant constraints. | Infrastructure decisions, long-term trade-offs |
| Persona | Style | Best For |
|---|---|---|
| Butterfield | "Every great product starts with a human problem." Watches users before asking. | User empathy, daily friction |
| Bezos | "Start with the customer and work backwards." Writes the press release first. | ROI, flywheel effects, one-way doors |
| Shreyas Doshi | "The most underrated PM skill is knowing what NOT to build." Priority frameworks, execution discipline. | Prioritization, execution rhythm |
| Persona | Style | Best For |
|---|---|---|
| Karpathy | "The most common error is not running the code." Builds first. | Prompt engineering, LLM behavior |
| Swyx | Sees everything as sense→think→act→observe loops. Paranoid about silent failures. | Agent architecture, file contracts |
| Simon Willison | "The best prompt is the one that ships." Datasette creator. | Practical prompt testing, cross-model compatibility |
| Persona | Style | Best For |
|---|---|---|
| Linus | "Talk is cheap. Show me the code." Finds the bug on line 47. | Bug hunting, race conditions |
| Carmack | Reviews from first principles. Hates unnecessary abstractions. | Performance, simplification |
| Persona | Style | Best For |
|---|---|---|
| Dan Abramov | "The mental model matters more than the API." React core, state philosophy. | State bugs, stale closures, component design |
| Evan You | "A framework should work where you can't feel it." Vue/Vite creator. | Build perf, DX, eliminating boilerplate |
| Guillermo Rauch | "If it doesn't deploy in one command, it's not done." Next.js/Vercel. | Deploy friction, framework selection, one-prompt-to-production |
| Persona | Style | Best For |
|---|---|---|
| DHH | "Convention over configuration." Hates microservices. | API design, monolith vs distributed, pragmatism |
| Antirez | "Simplicity is a prerequisite for reliability." Redis creator. | Data structures, concurrency, crash safety |
| Persona | Style | Best For |
|---|---|---|
| Miyamoto | Designs for the smile in the first 10 seconds. Subtraction design. | Onboarding, intuitive UX |
| Jonathan Blow | "Solve by understanding, not by trying everything." Braid, The Witness. | Deep puzzles, conceptual depth |
| Nicky Case | "Let people play with complex ideas." Evolution of Trust. | Web-native interactives, education |
| Persona | Style | Best For |
|---|---|---|
| Schneier | "Security is not a product, but a process." Thinks in threat models. | Threat modeling, trust boundaries |
| Mitnick | Thinks like an attacker. Humans are the weakest link. | Penetration testing, social engineering |
| Persona | Style | Best For |
|---|---|---|
| Don Norman | Father of UX. Affordances, signifiers, mental models. | Information architecture, confusion |
| Dieter Rams | "Good design is as little design as possible." | Visual cleanup, removing clutter |
| Jony Ive | "True simplicity is derived from so much more than just the absence of clutter." | Visual polish, premium feel, screenshot-worthy design |
| Persona | Style | Best For |
|---|---|---|
| Ben Treynor | "Hope is not a strategy." Invented SRE at Google. | Failure modes, reliability |
| Kelsey Hightower | Makes complex infra feel simple. "Do you need this?" | Simplification, build vs buy |
| Persona | Style | Best For |
|---|---|---|
| Martin Fowler | "Good programmers write code humans can understand." Refactoring father. | Separation of concerns, module boundaries |
| Leslie Lamport | "A distributed system is one where a computer you didn't know existed can break yours." | Consistency, ordering, concurrency |
| Persona | Style | Best For |
|---|---|---|
| Philip Tetlock | "Superforecasters" — who predicts well and why. | Human judgment, forecaster training |
| Nassim Taleb | "What happens on the worst day?" Black Swan, Antifragile. | Tail risk, position sizing, stress testing |
| Jim Simons | Renaissance Technologies. 66% annual returns for 30 years. | Signal extraction, feature engineering |
| Persona | Style | Best For |
|---|---|---|
| Paul Graham | "Write like you talk." One surprising insight per essay. | Article structure, core argument |
| Eugene Wei | "Status as a service." Understands algorithmic feeds. | Distribution mechanics, virality |
| Pieter Levels | "Ship fast, measure, iterate." 12 startups in 12 months. | MVP launch, growth hacking, distribution channels |
| Persona | Style | Best For |
|---|---|---|
| Kent Beck | "I'm not a great programmer; I'm just a good programmer with great habits." TDD pioneer. | Test-driven design, test architecture |
| James Whittaker | "Testing is not about finding bugs. It's about reducing risk." Google test director. | Test strategy, E2E planning |
| Michael Bolton | "Testing is not about proving the software works." Exploratory testing. | Edge cases, unexpected paths |
## Roundtable: [Topic]
### Participants
- Jensen (CEO) — "Double down on the GPU path. Ship it Tuesday."
- Linus (Reviewer) — "Race condition on line 47. Fix before anything."
### Consensus
- Everyone agrees: the feedback loop must close visibly
### Disagreements
- Jensen says accelerate shipping. Miyamoto says the player hasn't smiled yet.
### Actions
| Priority | Action | Raised By |
|----------|--------|-----------|
| P0 | Fix race condition | Linus |
| P1 | Ship feedback visibility | Jensen + Miyamoto || Feature | Other tools | Roundtable |
|---|---|---|
| Perspectives | Same-domain (multiple code reviewers) | Cross-discipline (CEO + Security + Game Designer) |
| Personas | Generic ("agent-1", "security-sentinel") | Named characters with distinct thinking styles |
| Variety | Same agents every time | Random selection from pool — same question, different insights |
| Integration | Framework / config required | One-sentence trigger via Claude Code skill |
| Output | Merged consensus | Preserved disagreements — conflicts are the most valuable part |
Edit references/personas.md:
### Your Expert · The Title
> "Their signature quote."
Personality description. What they obsess over. How they think.
**Looks for**: What they notice that others miss
**Best when**: When to pick this persona over alternativesEdit the role table in SKILL.md to add new roles or change when they trigger.
Feature planning:
"roundtable: we want to add a notification system. discuss." → Patrick evaluates long-term trade-offs, Architect designs pub/sub, Security checks for spam vectors
Post-implementation review:
"roundtable: review the changes I just made to the auth module" → Linus finds the edge case, Schneier traces the trust boundary, Kent Beck asks about test coverage
Architecture decision:
"the team should discuss whether to use WebSockets or SSE" → Lamport evaluates consistency, Kelsey Hightower asks "do you need this?", Carmack checks perf
Game design:
"roundtable: evaluate the player feedback system" → Miyamoto focuses on intuitive interaction, Nicky Case on interactive explanation, Karpathy on AI-driven adaptation
This skill was born from building Hermes Quest — a gamified RPG dashboard for an AI agent. During development, we ran roundtable discussions with 5+ expert personas to evaluate features, review code, and plan architecture. The pattern worked so well that we extracted it into a reusable skill.
Key insight: a PM, a game designer, and a code reviewer looking at the same codebase find completely different categories of problems. The PM found that the feedback loop was "a promise, not an experience." The code reviewer found a race condition in the timer cleanup. The game designer found that the core emotional beat was invisible. None of them would have caught what the others found.
Multi-agent debate is an active research area — it improves reasoning and reduces hallucination. Roundtable brings this to everyday development with named personas that produce genuinely different analysis angles.
- Claude Code (CLI) with background agent support
- Works with any Claude model (Opus recommended for best persona adherence)
MIT — Use it, fork it, add your own experts.
The most valuable insight is where the experts disagree.