Pilot rides the whole lifecycle — you just describe the work in your own words. This page shows what to say at each SDLC phase, which skill pilot routes to, which gates protect you, and what you get out.
tdd, diagnose, graphify) is an override, not a requirement.pilot off skips the next gate fire, /pilot-status shows what's armed, /pilot-doctor diagnoses the install.Three ways to drive. They compose — start hands-off, drop to manual when a phase needs your judgment.
"add rate limiting to the API" → pilot detects the phase from your words + project state and invokes the right skill. You approve plans, review diffs, say "go" between phases.
"use context7 for the docs, plan with writing-plans, tdd it, verify with playwright" → every named tool becomes a phase in a chain, executed in order. No keyword scoring.
/pilot-autopilot "<requirement>" → the whole loop runs unattended and stops exactly twice: plan approval and ship approval. Reply approved — your prompt is machine-witnessed.
Each card: what you type, where pilot routes, which gates are live, what artifact you walk away with.
init → gsd-new-project (structured, multi-session) — fires automatically when there's no CLAUDE.md.planning/ project state, CI + pre-commit quality-floor templates ready to activate with /pilot-flooroffice-hours (product interrogation); code intent → grill-with-docs → to-prd. Ambiguities become one focused question, not three assumptionssuperpowers:writing-plans (single-session) or gsd-plan-phase (multi-day); ceo-review to stress the plan strategically.pilot/acceptance.md (AC-001… checkboxes) — the ledger verify-gate later refuses to close uncheckedtdd (red-green-refactor, one commit per tracer slice); UI → ui-ux-pro-max / frontend-design; unfamiliar library → context7 pulls current docs firstverify runs the suite so the result is captured; UI changes get driven for real — playwright-cli first (open → snapshot → click → eval → screenshot), playwright MCP fallbackdiagnose — reproduce → hypothesize → instrument → fix → regression-test. Hypothesis before edit is non-skippable; no sleep-based band-aidssuperpowers:requesting-code-review (independent read); auth/crypto/network diffs trigger security-review — mandatory before Ship; github MCP reads real PR/CI stategsd-ship (PR → review → merge); release artifacts → version-bump (semver + changelog + tag); prod-bound → pre-deploy-checklist fires first, post-deploy-monitor watches aftermigration-safety (MIGRATION-SAFETY.md before pre-deploy); smells → improve-codebase-architecture; graphify maps the codebase into a queryable knowledge graph — and re-maps each cycle's delta at Capturecontinue to resume. Branch-scoped, so teammates run parallel cycles. autopilot off aborts. Approvals only count from your actual prompt (or /pilot-approve) — the model can't approve itselfThe habits that make pilot work across more than one keyboard.
bash dev/bootstrap-team.sh --check installs/verifies the pinned skill constellation, then /pilot-doctor. Ramp with docs/team-onboarding.md — ten minutes.
.pilot.json sets style + strictness + your test runner (test_patterns) so verify-gate recognizes real runs. Nobody inherits another dev's communication register.
Turn on team.shared_outcomes; dev/outcome-report.sh shows the team's first-pass-verified rate per user. High block counts = "done" is being claimed without tests.
Run the A/B protocol in docs/ab-method.md — pilot-on vs pilot-off weeks, binary scoring. Keep gates that earn their keep; cut the ones that don't.