SuperAPI AI Directory

si

General & other

View repository

Updated 10-15 21:47

About this skill

/si — Superintelligence Router

Route ANY ask through this 3-step protocol:

Step 1 — Classify intent

Parse $ARGUMENTS for verb. Match against these classes:

Execution verbs (route to: reasoning + swarm)

build, ship, deploy, fix, elevate, create, refactor, migrate, scaffold, wire, sync, audit-and-fix, redesign, rebuild, generate, render

Reasoning verbs (route to: reasoning only)

think, plan, compare, analyze, should, what if, why, evaluate, audit, research, explore, decide

Bare topic (no verb)

Route to: reasoning only, then ASK "Should I dispatch a swarm to execute this?"

Step 2 — Invoke /superintelligence

Always. Pass $ARGUMENTS as the topic. Capture the reasoning output.

Cost: ~5K input + ~3K output tokens, ~$0.05. Cheap enough to always run.

Step 3 — Route based on Step 1 classification

If execution intent

After /superintelligence returns, dispatch the appropriate swarm:

  • "build / scaffold / create / generate" → BUILD lane (e.g., /v BUILD or feature-dev or frankx-website-builder agent)
  • "deploy / ship / sync" → DEPLOY lane (e.g., /v DEPLOY)
  • "fix / refactor / migrate" → FIX lane
  • "elevate / redesign / rebuild" → ELEVATE lane (e.g., /v ELEVATE)
  • "audit-and-fix" → /v ELEVATE
  • Multi-domain ambiguity → ask which lane

If reasoning intent

Return /superintelligence verdict directly. Surface a one-line "Want to execute this?" footer.

If bare topic

Return /superintelligence verdict. Surface explicit "Should I dispatch a swarm to execute this?" question with 3 options (yes-execute / no-thinking-only / scope-first).

Cost discipline

  • Never auto-dispatch a swarm on a bare-topic ask
  • Never skip /superintelligence on execution intent (the framing is the value)
  • Per-invocation cost budget: $0.05 reasoning + variable swarm cost ($0.30-$3 per agent × N agents)
  • Compare to: a pure /v ELEVATE run is ~$1.50-$5 — /si adds $0.05 of reasoning at the front for better orchestration

Anti-patterns

  • Auto-dispatching swarm on every /si call (burns money on tasks that needed only thinking)
  • Skipping /superintelligence on execution (loses multi-perspective framing)
  • Treating /si as an alias for /superintelligence (it's a router, not a duplicate)

Cross-references

  • [[feedback_si_router_design]] — the design memory
  • /superintelligence skill — the reasoning primitive
  • /v skill — the Vercel/Next.js excellence swarm
  • /ultrawork — parallel agent execution
  • /ao — model routing