End-to-end feature pipeline with ONE human gate before code — Research (grilling rounds + parallel code scouts; a raw concept is stress-tested before convergence), then autonomous Plan → Implement → Review — top-tier planning over mapped context, cheap workers in parallel waves, a verifier per wave, an independent…
Spec-driven feature work in 4 phases — Specify, Design, Tasks, Execute — sized by complexity; atomic tasks and commits, Execute delegated to workers in waves plus an independent Verifier. Triggers on "specify/discuss feature", "design", "tasks", "implement", "validate", "verify work", "UAT", "record decision"…
Locates code in this repo without polluting the main context — where a symbol is defined, who consumes a route/component, where a module's rule lives, what's in a large file, the map of a module or feature. Use instead of grepping around. Not for editing, reviewing or deciding architecture. Always pass model — haiku…
Runs a terminal command in this repo (test, typecheck, lint, build, script) without dumping the output into the main context — full log to a file, returns exit code + the literal failures. For the main window's gates: the orchestrator's Build gate per wave and the Verifier's Final gate; a spec-worker runs its own…
Independent verifier for a ca-spec-driven feature, dispatched by the orchestrator after the last wave (author ≠ verifier) — spec-anchored coverage with evidence-or-zero, the Final gate (only full-suite run), mutation sensor, writes .specs/features/ /validation.md, returns the compact verdict. Never fixes code. Pass…
Implements ONE cluster of tasks from a ca-spec-driven feature (Execute) — scoped gate, one atomic commit per task, compact summary back. Dispatched only by the orchestrator, in parallel with other workers on the same checkout. Pass model per cluster — sonnet by default; haiku for pure mechanics; opus for domain…
Write a crash course — a study document, like a short documentary — about a concept, either one registered as debt in .learning/DEBT.md (e.g. "/crash-course DEBT-007") or any topic the user names ("/crash-course event sourcing", "explain properly what this backoff you added actually is", "make me study material on…
Register technical or cognitive debt about something AI implemented in the project, as a personal pending item in .learning/DEBT.md. Use whenever the user explicitly asks to record, log, note or flag a debt — "/debt", "log this as debt", "note this as tech debt", "I don't understand what you did here, flag it for me…
Expert, skeptical business analyst for a business / startup / product idea. Grills the user about the idea (one round of pointed questions at a time), then runs deep market research in the home country (default Brazil) and internationally — TAM/SAM/SOM, competitors, demand evidence, pricing, regulation, costs — builds…