the command router
one impeccable skill, 23 commands beneath it, routed by context.
what it teaches
- one entry point, many operations.
/menu pollution in AI coding tools worsens with every installed plugin. one skill plus a commands table beats twenty skills competing for the trigger. - a command maps to a reference file. each sub-command (
critique,audit,polish,bolder,quieter,distill,harden,live, …) owns one loaded-on-demand reference. the skill body stays a router; depth loads only when invoked. - routing is context-aware, never auto-run. with no argument, the tool reads project signals (design doc? recent critique score? dirty files? dev server running?) and recommends two or three commands with one-line reasons — menu is the fallback, recommendation is the lede, nothing runs without confirmation.
- maintenance tooling stays out of the design menu.
doctor,hooks,pinfollow a utility pattern, excluded from the command count. - discovery text is engineered. the skill description names trigger conditions, surfaces, symptoms — same description-trigger discipline as superpowers skill authoring.
why
twenty-three separate skills compete for attention in every session's skill discovery — the user sees /audit-ui, /fix-contrast, /check-motion, and twenty more, and the cost of choosing grows with each. one router loads its heavy reference only when a command fires. context-awareness turns "here's everything we can do" into "here's what your project needs next".
when to use
any tool family growing past a handful of operations: consolidate behind one router with on-demand references.
good and bad
bad
/audit-ui /fix-contrast /check-motion /review-copy /inspect-layout
many overlapping entry points make the right action harder to discover: a scan-and-guess tax on every invocation, wrong picks half the time, and each new menu item taxes every future session. heavy references loaded by trigger also burn context on sessions that never needed them.
good
/impeccable audit/impeccable polish/impeccable critique
one router exposes context-specific operations without making each one a global menu item. discovery cost stays flat as operations grow, reference material loads only when invoked, and context-aware recommendations turn the menu into a pointed answer.