Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Rijul's software-engineering brain trust **and decision-making partner**. 70 named native personas across 11 cells — drawn from the people who built the cloud, the databases, the languages, the runtimes, the web platform, the security and reliability disciplines, the DevOps movement, and the AI-codi
.claude/skills/coco-research-engineering-super-intelligence-team-skill-entry/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 33% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 17% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 23% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 414% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 139% | 0% |
Your software-engineering brain trust and decision-making partner. 70 named native personas across 11 cells — drawn from the people who built the cloud, the databases, the languages, the runtimes, the web platform, the security and reliability disciplines, the DevOps movement, and the AI-coding frontier — plus 9 personas cross-listed from the AI Super Intelligence Team. Reusable across any CoCo-routed prompt, invoked by the /SI-Eng-* slash commands. The team's primary purpose is to help take engineering decisions, not just review work after the fact.
> Status: Roster build complete (2026-05-30). > - Roster: 70 native personas across 11 cells, locked after a /ultra-think pass on the draft 65 (added Lamport, Lattner, Torvalds, Liskov, Perlman; reclassified Ghemawat as archetype; split the former devops-platform-ai-coding mega-cell into devops-platform + ai-assisted-coding). > - Cross-listed: 9 ML-systems / AI-coding voices carry teams: [ai-super-intelligence, engineering-super-intelligence] — single file, dual membership, no duplication. > - Commands: 25 /SI-Eng-* slash commands installed at ~/.claude/commands/, generated from the shared template by superintelligence/ai/scripts/build_commands.py --team Eng.
This file is the user-facing entry point. The machine source of truth is registry.json, regenerated from persona frontmatter by python3 superintelligence/engineering/scripts/build_registry.py.
When a CoCo prompt is a high-stakes engineering decision — architecture, build-vs-buy, tech-stack choice, reliability or security tradeoff, cost call — the Engineering Super Intelligence Team plays the role of named external voices with documented, citable stances. Instead of "the panel said," every claim is attributed to a specific engineer. This lets convene synthesis:
public_stance in every persona carries an evidence_url).The roster leans toward strong, opinionated, publicly-documented engineers — which makes it powerful for decide / tradeoff / stress-test / roast, and carries a known simplicity-maximalist / anti-hype tilt (Hickey, Carmack, DHH, Cantrill) that the architecture-and-process voices (Fowler, Hohpe, Newman, Kim, Forsgren) counterweight.
| Cell | Count | Focus | File | |---|---|---|---| | cloud-architecture | 8 | Cloud-scale system design, infra primitives, build-vs-managed | cells/cloud-architecture.md | | reliability-sre-obs | 7 | SRE practice, observability, incident response, resilience | cells/reliability-sre-obs.md | | data-and-storage | 8 | Databases, distributed data, consistency, distributed-systems theory | cells/data-and-storage.md | | security | 6 | Security architecture, cryptography, vuln research, disclosure policy | cells/security.md | | finops-cost | 4 | Cloud cost engineering, FinOps practice | cells/finops-cost.md | | languages-runtimes | 8 | Language design, type systems, compilers, runtimes | cells/languages-runtimes.md | | systems-programming | 7 | Low-level, OS, performance, systems craft | cells/systems-programming.md | | web-and-frontend | 6 | Frontend frameworks, web platform, UI engineering | cells/web-and-frontend.md | | architecture-testing-craft | 8 | Software architecture, DDD, testing discipline, craft | cells/architecture-testing-craft.md | | devops-platform | 6 | DevOps movement, platform engineering, internal developer platforms | cells/devops-platform.md | | ai-assisted-coding | 2 (+2 cross-listed) | Agentic dev tools, codegen, the AI-coding frontier | cells/ai-assisted-coding.md |
Listed by cell. Each has a YAML-frontmatter profile under personas/<slug>.md plus a research dump under research/<slug>/. Personas marked (archetype) use persistent_signals rather than recent signals (foundational figures or deliberately low-public-footprint).
Cloud Architecture (8): james-hamilton · werner-vogels · adrian-cockcroft · marc-brooker · brendan-burns · eric-brewer · colm-maccarthaigh · radia-perlman
Reliability, SRE, Observability (7): ben-treynor-sloss · betsy-beyer (archetype) · charity-majors · cindy-sridharan (archetype) · liz-fong-jones · nora-jones · tammy-butow (archetype)
Data and Storage (8): martin-kleppmann · jeff-dean · sanjay-ghemawat (archetype) · pat-helland · michael-stonebraker · andy-pavlo · joe-hellerstein · leslie-lamport
Security (6): bruce-schneier · alex-stamos · window-snyder · matthew-green · tavis-ormandy · katie-moussouris
FinOps and Cost (4): corey-quinn · jr-storment · mike-fuller · erik-peterson
Languages and Runtimes (8): guido-van-rossum · anders-hejlsberg · rich-hickey · graydon-hoare · brendan-eich · yukihiro-matsumoto · bjarne-stroustrup · chris-lattner
Systems Programming (7): john-carmack · bryan-cantrill · jonathan-blow · mitchell-hashimoto · ryan-dahl · brian-kernighan (archetype) · linus-torvalds
Web and Frontend (6): evan-you · dan-abramov · rich-harris · guillermo-rauch · ryan-carniato · adam-wathan
Architecture, Testing, Craft (8): martin-fowler · kent-beck · eric-evans · sam-newman · michael-feathers · dhh · gregor-hohpe · barbara-liskov (archetype)
DevOps and Platform (6): gene-kim · jez-humble · nicole-forsgren · kelsey-hightower · matthew-skelton · solomon-hykes
AI-Assisted Coding (2 native + 2 cross-listed): michael-truell · nat-friedman · andrej-karpathy (cross-listed from AI) · sasha-rush (cross-listed from AI)
These carry teams: [ai-super-intelligence, engineering-super-intelligence] and home_team: ai-super-intelligence. Their files live under superintelligence/ai/personas/; the Engineering registry references them via the cross_listed_from_ai field. No duplication.
andrej-karpathy · sasha-rush · tri-dao · bryan-catanzaro · andrew-feldman · albert-gu · horace-he · woosuk-kwon · tim-dettmers
superintelligence/engineering/
├── SKILL.md This file — user-facing entry.
├── ROSTER.md Locked roster ground-truth + build-wave manifest.
├── registry.json Machine source of truth. Read by slash commands.
├── personas/ 70 *.md files, one per native persona. YAML frontmatter + narrative sections.
├── cells/ 11 *.md cell summaries (generated by build_cells.py).
├── research/ 70 directories, one per persona. Raw research dumps.
└── scripts/
├── build_registry.py Regenerates registry.json from persona frontmatter.
└── build_cells.py Regenerates the 11 cell docs from registry + frontmatter.Templates (persona.md, convene.md) are shared one level up at superintelligence/templates/.
25 /SI-Eng-* command files live at ~/.claude/commands/, generated from the shared template by python3 superintelligence/ai/scripts/build_commands.py --team Eng. Architecture is orchestrator-first: every action verb invokes /SI-Eng-Orchestrate to pick a custom 16–32 persona team and gate on user approval before executing.
/SI-Eng — no args → roster + cell heatmap; with a subcommand, routes; with free text, defaults to :meeting./SI-Eng-Orchestrate "<prompt>" — scores all 70 personas (domain match 40% + cell coverage 30% + productive-conflict pairing 30%), picks 16–32, approval gate via AskUserQuestion, hard 16–32 band, re-picks every invocation./SI-Eng-Ask <slug> "<question>" — one persona in voice./SI-Eng-Huddle <cell-slug> "<topic>" — whole cell synthesizes./SI-Eng-Meeting "<prompt>" — full convene with mandatory attribution./SI-Eng-Read <slug> — print persona file inline./SI-Eng-Recruit <domain> "<why>" — propose new persona candidates for an under-covered domain./SI-Eng-Analyse · /SI-Eng-Decide (primary) · /SI-Eng-Review · /SI-Eng-Re-Analyse · /SI-Eng-Pre-Mortem · /SI-Eng-Post-Mortem · /SI-Eng-Full-Cycle · /SI-Eng-Tradeoff · /SI-Eng-Plan · /SI-Eng-Design · /SI-Eng-Vote · /SI-Eng-Debug · /SI-Eng-Stress-Test · /SI-Eng-Defend · /SI-Eng-Roast
/SI-Eng-Refresh · /SI-Eng-Verify · /SI-Eng-VoiceCheck--no-orchestrate — skip orchestrator; use all 70 personas.--cells <comma-list> — manually scope to cells.--personas <comma-list> — manually scope to slugs.superintelligence/templates/persona.md. Edit there first.python3 superintelligence/engineering/scripts/build_registry.py after any persona edit; then build_cells.py for the cell docs.public_stance has an evidence_url. No uncited claims.teams: [...] array and a home_team pointer. Never duplicate a persona file across teams./ultra-think before build: 5 canon adds, Ghemawat → archetype, mega-cell split, two all-male cells de-skewed (Liskov → craft, Perlman → cloud-architecture).| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 44,607 | 31,707 | -29% | 1 | 1 | 0% | 5,784 | 7,719 | +33% | 0 | 0 | — |
case-02 | fail→pass | 54,606 | 36,490 | -33% | 1 | 1 | 0% | 7,563 | 8,867 | +17% | 0 | 0 | — |
case-03 | fail→pass | 57,307 | 38,651 | -33% | 1 | 1 | 0% | 8,260 | 10,153 | +23% | 0 | 0 | — |
case-04 | pass→pass | 19,111 | 17,199 | -10% | 1 | 1 | 0% | 2,123 | 6,523 | +207% | 0 | 0 | — |
case-05 | pass→pass | 10,552 | 7,317 | -31% | 1 | 1 | 0% | 1,126 | 5,136 | +356% | 0 | 0 | — |
case-06 | pass→pass | 11,356 | 9,931 | -13% | 1 | 1 | 0% | 1,273 | 4,735 | +272% | 0 | 0 | — |
case-07 | fail→pass | 39,048 | 16,149 | -59% | 1 | 1 | 0% | 1,140 | 5,857 | +414% | 0 | 0 | — |
case-08 | fail→pass | 13,192 | 13,681 | +4% | 1 | 1 | 0% | 2,366 | 5,655 | +139% | 0 | 0 | — |
case-09 | fail→pass | 32,620 | 2,986 | -91% | 1 | 1 | 0% | 818 | 4,469 | +446% | 0 | 0 | — |
case-10 | fail→pass | 17,262 | 13,621 | -21% | 1 | 1 | 0% | 2,698 | 5,453 | +102% | 0 | 0 | — |
case-11 | fail→pass | 23,075 | 7,339 | -68% | 1 | 1 | 0% | 1,522 | 4,408 | +190% | 0 | 0 | — |
case-12 | fail→pass | 15,182 | 6,686 | -56% | 1 | 1 | 0% | 1,662 | 4,246 | +155% | 0 | 0 | — |
case-13 | fail→pass | 15,365 | 7,630 | -50% | 1 | 1 | 0% | 1,664 | 4,452 | +168% | 0 | 0 | — |
case-14 | fail→pass | 13,833 | 3,956 | -71% | 1 | 1 | 0% | 1,526 | 4,696 | +208% | 0 | 0 | — |
case-15 | fail→pass | 13,628 | 7,041 | -48% | 1 | 1 | 0% | 2,182 | 4,371 | +100% | 0 | 0 | — |
case-16 | fail→pass | 13,477 | 7,098 | -47% | 1 | 1 | 0% | 1,811 | 4,265 | +136% | 0 | 0 | — |
case-17 | fail→pass | 20,651 | 7,604 | -63% | 1 | 1 | 0% | 2,498 | 4,400 | +76% | 0 | 0 | — |
case-18 | fail→pass | 10,057 | 3,555 | -65% | 1 | 1 | 0% | 1,298 | 4,438 | +242% | 0 | 0 | — |
case-19 | fail→pass | 15,586 | 2,182 | -86% | 1 | 1 | 0% | 1,701 | 4,229 | +149% | 0 | 0 | — |
case-20 | fail→pass | 16,275 | 9,140 | -44% | 1 | 1 | 0% | 1,911 | 4,587 | +140% | 0 | 0 | — |
case-21 | fail→pass | 10,348 | 7,830 | -24% | 1 | 1 | 0% | 1,685 | 4,415 | +162% | 0 | 0 | — |
case-22 | fail→pass | 13,135 | 6,506 | -50% | 1 | 1 | 0% | 1,019 | 4,227 | +315% | 0 | 0 | — |
case-23 | fail→pass | 15,613 | 7,478 | -52% | 1 | 1 | 0% | 1,482 | 4,311 | +191% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 21 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +87 percentage points is the difference between those two pass rates over the 21 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.