Skip to content

superintelligence

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

superintelligence

README.md

Super Intelligence Teams

Note: Personas reference real public figures. Profiles are synthesized from public sources, with cited evidence — but roles and facts can change or go out of date, and quoted phrasing may be paraphrased. Treat them as illustrative expert lenses, not official statements. Not affiliated with or endorsed by the named individuals. See DISCLAIMER.md.

Reusable rosters of real-world thought-leader personas that act as parallel review-and-decision panels inside the CoCo Platform. Each team is a set of named experts with documented, citable public stances; you summon a custom 16–32 person panel for a prompt and every line of output is attributed to a specific person — never "the panel said."

The teams are a decision-making partner, not just a review surface. They are invoked through /SI-* slash commands installed at ~/.claude/commands/.

Teams

Team Command prefix Personas Cells Validated
AI Super Intelligence /SI-AI-* 59 8 built
Engineering Super Intelligence /SI-Eng-* 70 (+9 cross-listed from AI) 11 built
Product & Design Super Intelligence /SI-PD-* 56 (+1 cross-listed from Eng) 8 built
Finance Super Intelligence /SI-Fin-* 39 7 33/47 PASS
Trading Super Intelligence /SI-Trade-* 43 8 33/46 PASS
Risk & Compliance (GRC) /SI-GRC-* 28 (+12 cross-listed) 7 27/30 PASS
Strategy Super Intelligence /SI-Strat-* 23 (+2 cross-listed from P&D) 7 21/29 PASS
Data & Analytics Super Intelligence /SI-Data-* 25 7 24/29 PASS
Sales, GTM & Marketing /SI-GTM-* 20 (+3 cross-listed) 7 17/23 PASS

All 9 teams built and activated. The first three (AI, Engineering, Product & Design) were authored by Claude research agents and are built by construction. The six newer teams were built local-first then completed by Claude research agents (see Build pipeline below) and gated by the validator; the Validated column is the share that clears the strict bar (real cited live URLs, ≥2 signals or archetype persistent_signals, every stance cited). Personas below the bar are still real, grounded panel members flagged for incremental top-up — not omitted.

The machine-readable meta-registry is registry.json (kind: superintelligence-meta). Each team has its own <team>/registry.json, <team>/SKILL.md (human entry point), <team>/ROSTER.md (locked roster), personas/, cells/, research/, and scripts/.

Product + Design are one merged team (product-design), mirroring the earlier Cloud + Code → Engineering merge: discipline-pure cells preserve depth, while the team convenes both lenses by default. The same applies to Engineering, which merged cloud and coding concerns.

Architecture: orchestrator-first

Every action verb (decide, tradeoff, pre-mortem, …) calls the team's orchestrator first:

  1. Orchestrate — read <team>/registry.json, score all personas with 0.40·domain-match + 0.30·cell-coverage + 0.30·productive-conflict-pairing, greedily pick 16–32 personas (hard band), and present them for approval (confidence-tiered: high → auto-proceed, low → AskUserQuestion). Re-picks every invocation; no caching.
  2. Execute — the action verb consumes the approved roster and produces its verb-specific output (decision matrix, tradeoff table, ranked failure modes, …) with per-line attribution.

Identity verbs (ask, huddle, meeting, read) skip the orchestrator for explicit manual control. Global flags --no-orchestrate, --cells <list>, --personas <list> bypass or scope selection on any command.

Command surface (25 per team)

<prefix> (dispatcher) · -Orchestrate · identity: -Ask -Huddle -Meeting -Read · roster: -Recruit · action (15): -Analyse -Decide -Review -Re-Analyse -Pre-Mortem -Post-Mortem -Full-Cycle -Tradeoff -Plan -Design -Vote -Debug -Stress-Test -Defend -Roast · maintenance: -Refresh -Verify -VoiceCheck.

The /SI-* command files live in ~/.claude/commands/ (user-global), outside this repo. The repo carries the generator plus all persona/registry data, so the commands can be regenerated anywhere.

Cross-team meta-orchestrator

The hardest decisions span domains ("ship an AI compliance product" → AI + GRC + Engineering + Product + Finance). The meta-orchestrator is a top-level router with no team segment that assembles one panel across teams:

Command Role
/SI Dispatcher — lists built teams + the cross-team verbs.
/SI-Orchestrate "<prompt>" Cross-team selection + approval gate; returns the merged roster.
/SI-<Verb> "<prompt>" The 15 action verbs at the cross-team level (/SI-Decide, /SI-Tradeoff, /SI-Pre-Mortem, /SI-Review, /SI-Stress-Test, /SI-Plan, /SI-Design, /SI-Analyse, /SI-Vote, /SI-Debug, /SI-Defend, /SI-Roast, /SI-Post-Mortem, /SI-Re-Analyse, /SI-Full-Cycle). Each calls /SI-Orchestrate first.

Stage-A routing is deterministic + local. scripts/meta_select.py embeds the prompt against each built team's profile + cell descriptions using a local nomic-embed model and ranks → the top 1–4 relevant teams + candidate cells + proportional weights. It never loads persona files — the light layer that avoids the "load everything" trap. /SI-Orchestrate then drills into only the selected teams, allocates the 16–32 panel by relevance weight, runs a cross-team tension pass, and attributes every line to persona + team. Auto-delegates to a single team's orchestrator when one domain clearly dominates.

# example: cross-team panel decides a cross-cutting question
/SI-Decide "Should we build and sell an AI-powered audit-automation product?"
#   -> meta_select routes to GRC + AI + Engineering + Finance + Product
#   -> ~24-persona panel, approval gate, decision matrix with cross-team dissent

Generated by scripts/build_meta_commands.py, which reads the meta-registry live — teams auto-join the cross-team surface the moment their built flag flips.

Build pipeline (how the six newer teams were built)

Persona generation was an explicit cost/quality experiment, settled by evidence:

  1. Local-first (build_local.py, qwen3.6 via LM Studio) + free retrieval (retriever.py: DuckDuckGo + trafilatura, deep + recency-filtered). ~$0, but low instruction-following yield.
  2. gpt-5.4-nano (build_nano.py, QB AI Gateway) — fast, ~$0.003/persona, 20-parallel; retrieval-capped.
  3. gemini-3.5-flash (build_cursor.py, Cursor headless agent) — top quality, but agent-latency-flaky at batch scale.
  4. Claude research agents (WebSearch) — the quality winner; built the pending tail via a fan-out workflow.

Plus cheap repair/recalibration: yaml_repair.py + topup_cheap.py (fix malformed YAML, prune dead URLs), and a recalibrated validator (*/scripts/validate_persona.py) that keeps the anti-fabrication core (≥4 live real URLs, every stance cited, no invented quotes) while relaxing arbitrary quotas and tolerating link-rot. activate_teams.py flips built=true + regenerates commands for teams clearing threshold. Full write-up in QUALITY-FINDINGS.md.

Persona schema

Every persona is a Markdown file with YAML frontmatter + narrative sections, conforming to templates/persona.md. Key invariants:

  • teams: [...] array + home_team — a persona can belong to multiple teams via a single file (no duplication). Cross-team members are referenced from the other team's registry by relative path.
  • Every public_stance carries an evidence_url. No uncited claims.
  • Validator (recalibrated): ≥4 live real URLs, ≥2 recent_signal_12mo (each with a real url), every public_stance cited. Foundational / low-public-footprint / deceased figures are status: archetype and use ≥2 persistent_signals instead. Dead/404 URLs are tolerated as link-rot (pruned, not failed).
  • productive_conflict_with wires the disagreements the orchestrator mines for tension (e.g. Hinton↔LeCun, DHH↔Fowler, Eyal↔Harris).

Regenerating

# Per-team commands (omit --team for all): AI | Eng | PD | Fin | Trade | GRC | Strat | Data | GTM
python3 superintelligence/ai/scripts/build_commands.py --team Fin

# Cross-team meta commands (/SI, /SI-Orchestrate, /SI-<Verb>) — reads the meta-registry live:
python3 superintelligence/scripts/build_meta_commands.py

# Registry + cell docs + SKILL for a team (run after editing any persona file):
python3 superintelligence/<team>/scripts/build_registry.py
python3 superintelligence/<team>/scripts/build_cells.py
python3 superintelligence/scripts/build_skill.py <team>

# Re-validate + (re)activate a team once its PASS rate clears threshold:
python3 superintelligence/<team>/scripts/validate_all.py
THRESHOLD=0.70 python3 superintelligence/scripts/activate_teams.py

build_commands.py is the single source of truth for the command surface; all teams are stamped from one shared template, so a fix there propagates to every team. Per-team specifics (cells, gap doc, counts) live in its TEAMS dict.

Conventions

  • Schema source of truth: templates/persona.md. Edit there first.
  • Attribution mandatory: every claim names a persona or a cell.
  • Documentation is always full English prose (no caveman compression) — applies to persona files, cell docs, SKILL files, and this README.
  • Cross-team personas: one file, teams: [...] + home_team; never duplicate.

Layout

superintelligence/
├── README.md                  This file.
├── registry.json              Meta-registry (9 teams, default_team, build status).
├── DISCLAIMER.md              Real-figure / illustrative-composite disclaimer.
├── QUALITY-FINDINGS.md        Build-pipeline experiment + validator recalibration write-up.
├── ACTIVATION-REPORT.md       Per-team PASS rates + activation decisions.
├── META-ORCHESTRATOR-DESIGN.md  Cross-team router design + locked decisions.
├── templates/                 Shared persona.md + convene.md.
├── scripts/                   Cross-team: meta_select.py, build_meta_commands.py,
│                              build_nano/build_cursor, yaml_repair, topup_cheap,
│                              activate_teams, build_skill, validate/registry generics.
├── ai/                        AI team (59 personas, 8 cells) + scripts.
├── engineering/               Engineering team (70, 11 cells) + scripts.
├── product-design/            Product & Design team (56, 8 cells) + scripts.
├── finance/                   Finance team (39, 7 cells) + scripts.
├── trading/                   Trading team (43, 8 cells) + scripts.
├── risk-compliance/           Risk & Compliance / GRC (28 +12 x-listed, 7 cells) + scripts.
├── strategy/                  Strategy team (23 +2 x-listed, 7 cells) + scripts.
├── data-analytics/            Data & Analytics team (25, 7 cells) + scripts.
└── gtm/                       Sales, GTM & Marketing team (20 +3 x-listed, 7 cells) + scripts.

Each team dir: registry.json, roster.json (locked), SKILL.md, personas/, cells/, research/ (gitignored), scripts/.