coco-research AI Super Intelligence Team — SKILL entry
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.
What this team is for
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:
- Surface real disagreement (Hinton vs LeCun on existential risk; Karpathy vs Schulman on RL).
- Anchor stances to citable evidence (every public_stance in every persona has an
evidence_url). - Stay honest about who actually drove which decision (lead-driver vs validator vs specialist vs swing).
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.
Cells (8)
| 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 |
Personas (59)
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
Files in this team
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.
Slash commands (Phase 3 — installed)
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.
Dispatcher (1)
| 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. |
Orchestrator (1)
| 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. |
Identity surface — explicit overrides (4)
| 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 |
Roster management (1)
| 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. |
Action surface (15) — orchestrator-first by default
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 |
Global flags
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 |
Conventions enforced by every command
- Registry is source of truth. Re-read every invocation; no caching.
- Attribution at every line. No "the team said" — name a persona or a cell.
- Approval gate mandatory for action verbs unless an explicit flag bypasses.
- Hard 16-32 size band on orchestrator output.
- Full English prose — no caveman compression in command output.
Cross-team note
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 strategy
All 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.
Conventions
- Schema source of truth:
templates/persona.md. Edit there first; persona files conform. - Registry regeneration: run
python3 superintelligenceTeam/scripts/build_registry.pyafter any persona edit. - Caveman mode does NOT apply to persona files, cell files, or this SKILL.md. Documentation is always full English prose per the project rule.
- Citation is mandatory. Every
public_stancehas anevidence_url. No uncited claims. - v2 panel preservation. Personas with
v2_panel_attributionentries anchor to actual material from the founding Memory v2 synthesis on 2026-05-26. Convene uses those first. - Confidence calibration.
confidence < 0.85means we suspect identifier or biographical detail may need re-verification. Persona is still usable but flag in convene.
Build provenance
- Forged 2026-05-27 during /coco session.
- Quality bar set by Karpathy reference (12 cited URLs, 5 recent_signal_12mo entries, full 6-section narrative).
- 46 remaining personas built by parallel research sub-agents in 6 waves (8 + 8 + 8 + 8 + 8 + 6).
- Several agent runs corrected the user-supplied hints with verified facts (e.g., Lilian Weng's PhD = Indiana not NYU; Sergey Levine's PhD advisor = Vladlen Koltun not Pieter Abbeel; Barret Zoph fired from TML Jan 2026 and returned to OpenAI; Aditya Ramesh stayed at OpenAI Worldsim VP not TML).
- All persona files written in full English prose. Caveman mode active in chat throughout.