Your AI research and engineering brain trust. 59 named personas across 8 cells covering frontier labs, applied product, model architecture, reasoning/RL/agents, alignment and interpretability, theory and science of DL, multimodal and…
インストール方法を見るインストールする前に、エージェントに与えられる指示の中身を確認できます。
Your AI research + engineering brain trust and decision-making partner. 59 named personas drawn from the world's frontier model labs (US, UK, China, France), top AI universities, the open-source ecosystem, the independent-evaluation and AI-governance ecosystem, and product-history archetypes. Reusable across any CoCo-routed prompt — invoked by /superintelligenceTeam-* slash commands. The team's primary purpose is to help take decisions, not just review work after the fact — every cell carries a decision-relevant lens that convene synthesis explicitly draws on.
Status: All three phases complete.
- Phase 1: 49 personas across 8 cells (baseline roster + founding v2 panelists).
- Phase 2: 10 priority additions from the EXPANSION.md gap analysis (roster grew to 59).
- Phase 3: 22 slash commands installed at
~/.claude/commands/superintelligenceTeam*.md— orchestrator-first action surface plus explicit-override identity surface. See Slash Commands section below.
This file is the user-facing entry point for the team. The machine source of truth is registry.json, regenerated from persona frontmatter by python3 superintelligenceTeam/scripts/build_registry.py.
When a CoCo prompt is high-stakes enough to deserve a parallel review panel, the AI Super Intelligence Team plays the role of named external voices. Instead of "the panel said," every claim is attributed to a specific researcher with documented stances. This lets convene synthesis:
evidence_url).The 5-cell × 4-persona founding Memory v2 panel from May 26-27, 2026 was the first consumer. This team supersedes that roster: five of those panelists (Karpathy, Wei, LeCun, Tri Dao, Hyung Won Chung) continue here with their v2_panel_attribution records preserved. The rest of the founding v2 panel (memory / cloud / security / privacy personas) belong to future Super Intelligence Teams (Memory Systems, Cloud, Data, Compliance) and are not part of this AI-focused roster.
| Cell | Count | Focus | File |
|---|---|---|---|
frontier-labs-research | 6 | CSO / co-founder tier at frontier labs (US, China, France) | cells/frontier-labs-research.md |
applied-ai-leadership | 7 | Product / strategy founders shipping AI (incl. archetypes) | cells/applied-ai-leadership.md |
model-architects | 7 | Pretraining + scaling + model design + retrieval + open-source | cells/model-architects.md |
reasoning-rl-agents | 7 | Post-training, RL, agentic systems, test-time compute | cells/reasoning-rl-agents.md |
alignment-interp-safety | 9 | Alignment + mech interp + safety policy + independent evals + governance | cells/alignment-interp-safety.md |
theory-science | 9 | DL theory + science + Turing laureates + AI-for-science + common-sense | cells/theory-science.md |
multimodal-embodied | 7 | Vision + diffusion + robotics + embodied | cells/multimodal-embodied.md |
systems-kernels-serving | 7 | Kernels + serving + GPU + quantization + anti-NVIDIA silicon | cells/systems-kernels-serving.md |
Listed by cell. Each has a YAML-frontmatter profile under personas/<slug>.md plus a research dump under research/<slug>/. Bold marks Phase 2 additions (built 2026-05-28 from EXPANSION.md).
Frontier Labs Research (6): ilya-sutskever · dario-amodei · demis-hassabis · jakub-pachocki · liang-wenfeng · arthur-mensch
Applied AI Leadership (7): sam-altman · mira-murati · greg-brockman · aravind-srinivas · aidan-gomez · elon-musk · steve-jobs (archetype, deceased 2011)
Model Architects (7): andrej-karpathy · jared-kaplan · noam-shazeer · jason-wei · sebastian-raschka · patrick-lewis · thomas-wolf
Reasoning, RL, Agents (7): john-schulman · noam-brown · hyung-won-chung · nathan-lambert · barret-zoph · karina-nguyen · sasha-rush
Alignment, Interp, Safety (9): chris-olah · paul-christiano · jan-leike · dan-hendrycks · stuart-russell · neel-nanda · lilian-weng · beth-barnes · helen-toner
Theory and Science (9): yann-lecun · yoshua-bengio · geoffrey-hinton · john-jumper · percy-liang · christopher-manning · yejin-choi · sara-hooker · aleksander-madry
Multimodal, Embodied (7): fei-fei-li · pieter-abbeel · sergey-levine · chelsea-finn · robin-rombach · aditya-ramesh · prafulla-dhariwal
Systems, Kernels, Serving (7): tri-dao · bryan-catanzaro · andrew-feldman · albert-gu · horace-he · woosuk-kwon · tim-dettmers
superintelligenceTeam/
├── SKILL.md This file — user-facing entry.
├── registry.json Machine source of truth. Read by slash commands.
├── EXPANSION.md (Phase 2) Gap-analysis and roster-expansion candidates.
├── templates/
│ ├── persona.md Schema source-of-truth for every persona file.
│ └── convene.md Multi-persona session template.
├── personas/ 59 *.md files, one per persona. YAML frontmatter + 6 narrative sections.
├── cells/ 8 *.md cell summaries.
├── research/ 59 directories, one per persona. Raw research dumps so future re-syntheses don't recrawl.
└── scripts/
└── build_registry.py Regenerates registry.json from persona frontmatter. Run after any persona edit.
<a id="slash-commands"></a>
22 slash command files live at ~/.claude/commands/superintelligenceTeam*.md (user-global install, parallel to the existing /team family). Each is auto-registered as a discoverable Skill — no separate SKILL.md registration is needed. The architecture is orchestrator-first: every action verb invokes the orchestrator to pick a custom 16-32 persona team and gate on user approval before executing.
| Command | Purpose |
|---|---|
/superintelligenceTeam | No args → print roster + cell heatmap. With a subcommand as first token, routes to the sibling file. With free text and no subcommand, defaults to :meeting. |
| Command | Purpose |
|---|---|
/superintelligenceTeam:orchestrate "<prompt>" | Standalone team selection. Reads registry.json, scores all 59 personas via domain match (40%) + cell coverage (30%) + productive-conflict pairing (30%), picks 16-32, asks for user approval via AskUserQuestion with per-persona one-line rationale. Hard 16-32 size enforcement. Re-picks every invocation. Does NOT load CoCo. |
| Command | Purpose | Skips orchestrator? |
|---|---|---|
/superintelligenceTeam:ask <slug> "<question>" | 1-on-1 with one persona in their voice. | Yes |
/superintelligenceTeam:huddle <cell-slug> "<topic>" | Whole cell (4-9 personas) synthesizes. | Yes |
/superintelligenceTeam:meeting "<prompt>" | Full 59-persona convene with mandatory attribution. | Yes |
/superintelligenceTeam:read <slug> | Print the persona file inline (not voice-channeled). | Yes |
| Command | Purpose |
|---|---|
/superintelligenceTeam:recruit <domain> "<why>" | Propose 2-3 new persona candidates for an under-covered domain. Reads EXPANSION.md to avoid re-proposing. Does not write personas itself. |
All 15 invoke /superintelligenceTeam:orchestrate first unless --no-orchestrate, --cells, or --personas flag is supplied.
| Command | Output shape |
|---|---|
/superintelligenceTeam:analyse "<topic>" | Per-persona analysis + synthesis table of strongest signals |
/superintelligenceTeam:decide "<question>" | Primary verb. Decision matrix: options × personas × verdict + recommendation + named dissent |
/superintelligenceTeam:review <target> | Multi-persona findings classified CRITICAL/MAJOR/MINOR/SUGGESTION + ship verdict |
/superintelligenceTeam:re-analyse "<topic>" [--prior <path>] [--evidence <text>] | Updated stances + diff vs prior analysis |
/superintelligenceTeam:pre-mortem "<plan>" | Ranked failure modes + early warning signs + mitigations |
/superintelligenceTeam:post-mortem "<what failed>" | Per-persona 5 Whys + most-likely root cause + remediation plan |
/superintelligenceTeam:full-cycle "<topic>" | Heaviest verb. Chains :review → :analyse → mitigate-risk → :decide → finalize as real subcommand invocations. 5× latency. Finalized action plan. |
/superintelligenceTeam:tradeoff "<A vs B>" | Side-by-side dimensions × options table + most-opposed cell named |
/superintelligenceTeam:plan "<goal>" | Phased plan with owner cell + dissenting voices per phase |
/superintelligenceTeam:design "<feature>" | Component-level architecture with per-decision attribution |
/superintelligenceTeam:vote "<binary question>" | Yes/no per persona + tally by cell + recommendation |
/superintelligenceTeam:debug "<problem>" | Ranked root-cause hypotheses + diagnostic test order |
/superintelligenceTeam:stress-test "<proposal>" | Adversarial attacks per persona + severity matrix + SHIP/HARDEN-THEN-SHIP/REFRAME/DO-NOT-SHIP verdict |
/superintelligenceTeam:defend "<position>" [as <slug>] | Steelman with sharpened claim + counter-objection rebuttal + "what would change my mind" |
/superintelligenceTeam:roast "<thing>" | One-line cutting roast per persona + single line that hurt most + convergent critique |
Every command accepts these:
| Flag | Effect |
|---|---|
--no-orchestrate | Skip orchestrator; use all 59 personas |
--cells <comma-list> | Manually scope to specific cells; skip orchestrator |
--personas <comma-list> | Manually scope to specific persona slugs; skip orchestrator |
This is the first Super Intelligence Team in CoCo. Planned follow-ons:
cloud-super-intelligence — AWS / GCP / Azure architecturefinance-super-intelligence — FP&A, accounting, finance opscoding-super-intelligence — software-engineering practicedesign-super-intelligence — UX, design systems, visual designproduct-super-intelligence — PM craft, product strategyAll teams share the same schema (teams: [...] array on each persona, functional cell slugs, slash-command surface). A persona may belong to multiple teams — Karpathy will likely also appear in a future coding-super-intelligence team because of his pedagogical reach into engineering practice. The teams: field is an array exactly so this is cheap.
templates/persona.md. Edit there first; persona files conform.python3 superintelligenceTeam/scripts/build_registry.py after any persona edit.public_stance has an evidence_url. No uncited claims.v2_panel_attribution entries anchor to actual material from the founding Memory v2 synthesis on 2026-05-26. Convene uses those first.confidence < 0.85 means we suspect identifier or biographical detail may need re-verification. Persona is still usable but flag in convene.まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Train and optimize AI agents using Microsoft's Agent Lightning framework with reinforcement learning. Use when setting up agent training, instrumenting agents with tracing, configuring LightningStore, implementing reward functions, or optimizing prompts with RL/APO algorithms.
日本語の概要は準備中です。原文の説明を表示しています。
Post-run self-evaluation system that scores agent output on correctness, clarity, actionability, and conciseness. Use after /team runs, skill executions, or when explicitly asked to evaluate output quality.
日本語の概要は準備中です。原文の説明を表示しています。
Create AI marketing videos for ads, promos, product launches, and brand content. Models: Veo, Seedance, Wan, FLUX for visuals, Kokoro for voiceover. Types: product demos, testimonials, explainers, social ads, brand videos. Use for: Facebook ads, YouTube ads, product launches, brand awareness. Triggers: marketing video, ad video, promo video, commercial, brand video, product video, explainer video, ad creative, video ad, facebook ad video, youtube ad, instagram ad, tiktok ad, promotional video, launch video
日本語の概要は準備中です。原文の説明を表示しています。
Use when building AI features into a product: LLM integration, RAG pipelines, guardrails, streaming, AI UX, prompt engineering, or AI cost control. Treats prompts as code and validates every model output.
日本語の概要は準備中です。原文の説明を表示しています。
Use when designing a new REST or GraphQL API, reviewing an API spec before implementation, setting team API standards, or migrating REST to GraphQL. Covers resources, HTTP semantics, pagination, error handling, and pitfalls.
日本語の概要は準備中です。原文の説明を表示しています。
Use for authorized security assessment of REST, GraphQL, WebSocket, or SOAP APIs, including discovery, authentication, authorization, rate-limit, and CI/CD testing.
日本語の概要は準備中です。原文の説明を表示しています。