Skip to content

Commands

CommandWhat it does
/atelier:bootstrapOne-shot setup: workspace, SQLite store, environment doctor, companion check. Once per project.
/atelier:clarifyUnderstand-first: reads everything, then asks only the true unknowns, batched once.
/atelier:planArchitect writes the plan and the done-state contract (measurable criteria) into the store.
/atelier:buildExecutes the plan through the loop engine. Refuses to finish with unmet criteria.
/atelier:reviewRuns the gates — ponytail, graph impact, lint/typecheck, design tokens, file sizes — and records evidence per criterion.
/atelier:shipFinal release gate: verifies the entire contract, produces the ship report.
/atelier:statusCurrent phase: active plan, met/unmet criteria, capabilities, recent activity.
/atelier:doctorHealth check: environment, store integrity, companion MCPs, graph freshness.
/atelier:learnIngests screenshots, URLs, code, repos → workspace overlays and registry rows.
/atelier:promotePromotes an overlay into the project’s git-tracked design docs.
/atelier:updatePulls the latest plugin guidance, migrates the store (idempotent, self-healing), verifies health.

Everything durable sits in workspace/atelier.db — plans, criteria, rules, registry, learning events — served as token-lean pipe rows, not JSON. A SessionStart hook surfaces the active plan and unmet criteria after any restart or compaction, so work resumes exactly where it stopped.