SI convenes a panel of 16–32 modeled experts to argue about your question and hand you an attributed verdict. There are 242 slash commands under the hood, but the system routes for you — you only need to remember one.
/SIThe front door. It infers which teams, which verb, and which dials fit your question, confirms in one line, and runs.
/SI "should we build the export feature or buy a vendor tool?"
The dials are inferred from your wording: “make them argue” turns on debate rounds, “in depth” turns on full detail, “be brutal” routes to Roast.
/SI "kill the v1 UI or keep maintaining both? make them argue it out"
Already know the domain? The team front doors work the same way — plain English in, verb and dials inferred.
/SI-PD "review this onboarding flow, be brutal"
/SI (or /SI-<Team>) followed by plain English. Everything below this point is how it works under the hood, plus power-user shortcuts for when you want exact control.
Every command follows the same pipeline: infer the verb → route to teams → select the panel → approve → run. The front door first maps your wording to a verb (an intent table: “A or B” → Decide, “tear this apart” → Stress-Test, and so on — if two genuinely fit, it asks one quick question). Team routing is done by a local embedding model that scores your prompt against each team's profile and specialist cells (no cloud call). Single-domain questions auto-delegate to that team; multi-domain questions get a cross-team panel with seats split proportionally to relevance.
Three details worth knowing:
| Mechanism | What it means for you |
|---|---|
| Productive conflict pairing | 30% of an expert's selection score comes from how well they clash with others already on the panel. The roster is built to disagree, not to echo. |
| Confidence-tiered approval | When the pick is obviously right you get a one-line summary and it proceeds. When it's ambiguous or high-stakes, you get an approve / edit / re-pick gate. |
| Stance validation | Every claim in the output traces to that expert's documented public stances (with source URL) or is labeled [extrapolated]. Over 25% extrapolated triggers a low-grounding warning. |
--debate flagBy default each expert speaks once and the verb synthesizes (fast, cheap). With --debate, the orchestrator runs the panel as reacting rounds before handing results to the verb. The round protocol lives in exactly one place — the orchestrator — so all 130 debate-capable verbs share the same mechanics.
| Mode | Cost | When | Example |
|---|---|---|---|
(no flag) | Cheapest — single pass | Default. Quick takes, routine reviews. | /SI-Decide "ship Friday or wait for the fix?" |
--debate | Moderate — rounds run in-session | Contested calls where positions should be tested against each other. | /SI-Decide --debate "sunset the v1 API or support both?" |
--debate=full | Expensive — one subagent per expert per round | High-stakes, irreversible decisions only. | /SI-Decide --debate=full "migrate all customers to the new platform?" |
# You never need the flag — saying it in the prompt works: /SI "keep building the custom dashboard or adopt the vendor one? make them argue it out" # Debate on each verb that supports it: /SI-Decide --debate "build the integration in-house or buy the connector?" /SI-Tradeoff --debate "monorepo vs polyrepo for a six-person team" /SI-Analyse --debate "what does on-device AI mean for our roadmap?" /SI-Review --debate "review docs/pricing-proposal.md" /SI-Pre-Mortem --debate "we migrate billing to the new provider next quarter" /SI-Post-Mortem --debate "the launch missed signup targets by 60%" /SI-Stress-Test --debate "our plan to deprecate the v1 API in 90 days" /SI-Plan --debate "roll out SSO to every enterprise customer this quarter" /SI-Design --debate "notification service fanning out to email, chat, and in-app" /SI-Debug --debate "p99 latency tripled after the deploy but CPU is flat" /SI-Defend --debate "we should stay on SQLite instead of moving to Postgres" /SI-Re-Analyse --debate --prior notes/pricing-analysis.md --evidence "competitor cut prices 30%" /SI-Full-Cycle --debate "evaluate and decide on the new vendor platform" # heaviest combo — it will warn first # Team-scoped versions take it too: /SI-PD-Decide --debate "infinite scroll or pagination for the activity feed?" /SI-Eng-Design --debate "event bus vs direct service calls for order processing" /SI-GRC-Stress-Test --debate "our data-retention policy before the audit"
Vote and Roast deliberately don't take the flag — a binary tally and a one-liner roast have nothing to debate.
Every command follows one pattern. Once you know it, you know all 242:
/SI[-Team][-Verb] "your prompt" [flags] # no team, no verb → /SI "..." routes everything for you # verb only → /SI-Decide "..." cross-team panel, that verb # team + verb → /SI-PD-Decide "..." one team's roster, that verb # team only → /SI-PD prints that team's roster
| Code | Team | Experts | Reach for it when the question is about… |
|---|---|---|---|
PD | Product & Design | 56 | product strategy, UX, design systems, user research, what to build/SI-PD "infinite scroll or pagination for the activity feed?" |
Eng | Engineering | 70 | architecture, code, infrastructure, reliability, security engineering/SI-Eng "event bus or direct service calls for order processing?" |
AI | AI | 59 | models, agents, ML systems, evals, AI safety and capability/SI-AI "fine-tune a small model or prompt a frontier one for triage?" |
Data | Data & Analytics | 29 | pipelines, metrics, experimentation, analytics engineering/SI-Data "is this A/B test result real or peeking bias?" |
Fin | Finance | 47 | budgets, unit economics, valuation, capital allocation/SI-Fin "does annual prepay discounting actually improve our cash position?" |
Trade | Trading | 46 | markets, risk models, derivatives, market microstructure/SI-Trade "stress-test our exposure if rates jump 100bps" |
GRC | Risk & Compliance | 30 | governance, privacy, regulatory exposure, audit, controls/SI-GRC "review our data-retention policy before the audit, be brutal" |
Strat | Strategy | 29 | competitive positioning, market entry, long-range bets/SI-Strat "enter the SMB segment or double down on enterprise?" |
GTM | Sales, GTM & Marketing | 23 | pricing, positioning, launch, pipeline, growth/SI-GTM "usage-based pricing or seats for the new tier?" |
Every card has two tabs — flip to Examples for a ready-to-paste command per verb.
Decide — pick between options, named dissentTradeoff — A-vs-B scored on 4–7 dimensionsVote — yes/no tally on a binary questionAnalyse — multi-lens read of a topicReview — critique a doc, PR, or designRe-Analyse — rerun a prior analysis with new evidenceDebug — ranked root-cause hypotheses + test orderPre-Mortem — what kills this in 12 monthsPost-Mortem — 5-whys root cause after a failureStress-Test — adversarial attack, no diplomacyRoast — one cutting line per expertDefend — steelman a position properlyPlan — phased plan with owners and risksDesign — architecture pass, per-decision attributionFull-Cycle — review → analyse → mitigate → decide (heavy, ~5× latency)Ask <slug> — 1-on-1 with one expert in their voiceHuddle <cell> — one specialist cell synthesizesMeeting — the entire team convenesRead <slug> — print an expert's profileOrchestrate — panel selection only (verbs call this for you)Recruit — propose new experts for a gapRefresh — update stale expert profilesVoiceCheck / Verify — quality checks on voices and citations| Flag | Effect | Example |
|---|---|---|
--deep | Full per-expert detail. Default output is the short tier: verdict + top dissents, ~10–20 lines. | /SI-Analyse --deep "our churn doubled in Q2"or just say “in depth” in the prompt |
--debate / --debate=full | Reacting rounds before synthesis (see above). | /SI-Decide --debate "kill or keep the v1 UI?"or just say “make them argue” |
--no-orchestrate | Skip selection; use the whole team. | /SI-PD-Review --no-orchestrate "the new onboarding flow" |
--cells a,b / --personas x,y | Hand-pick the scope; skips orchestration. | /SI-GRC-Analyse --cells privacy-data-protection "chat-log retention"/SI-Strat-Decide --personas michael-porter,roger-martin "enter the SMB market?" |
--coco-off | Use the standalone selector instead of CoCo's context-aware routing. | /SI-Decide --coco-off "A or B?" |
| You're thinking… | Type |
|---|---|
| “I don't know who should look at this” | /SI "…" |
| “A or B?” | /SI-Decide "A or B?" |
| “This is a big, contested call” | /SI-Decide --debate "…" |
| “Tear this plan apart before I commit” | /SI-Stress-Test "…" or /SI-Pre-Mortem "…" |
| “Review this PR / doc / design” | /SI-Eng-Review <path> (or PD/GRC flavor) |
| “Why did this fail?” | /SI-Post-Mortem "…" · “What's broken?” → /SI-Debug |
| “Plan this out” / “architect this” | /SI-Plan "…" / /SI-Design "…" |
| “What would Ann Cavoukian say?” | /SI-GRC-Ask ann-cavoukian "…" |
| “Who's even on these teams?” | /SI (no args) or /SI-PD for one team |
| “Be brutal” | /SI-Roast "…" |
Full structural audit of the live installation, with every finding fixed the same day. Verdict: the orchestrator is working, all known gaps closed.
| Check | Result |
|---|---|
| All 242 command files installed and loading | PASS |
| Every file path referenced by commands resolves | PASS |
Embedding router meta_select.py live-tested | PASS — correct team + cell scores returned |
| All 9 team registries consistent; cross-team listings resolve; every expert file on disk registered and selectable | PASS — 0 broken references, 0 orphans |
| Debate wiring: protocol in exactly 10 orchestrators, 117 team verbs + 13 cross verbs forward the flag, Vote/Roast clean | PASS — matches design exactly |
| Generators portable (env-overridable repo path, no dead-checkout references); team metadata derived live from registries | PASS |
| 27 experts (incl. Porter, Rumelt, Marks) were invisible to selection — root cause was broken YAML frontmatter; repaired deterministically and re-registered | FIXED 2026-06-11 |
ROSTER.md generated for all 6 missing teams (new build_roster_md.py, regenerable from registries) | FIXED 2026-06-11 |
validate_all_generic.py runs team-aware from the repo root; generator metadata drift eliminated | FIXED 2026-06-11 |