Commands reference

Setup commands

Command	What it does
/kickoff	Interactive project setup — asks 5 questions, populates CLAUDE.md, .claude/personas.md, optionally seeds GitLab issues
/import-spec	Convert a PRD, spec, or feature list into structured GitLab issues. Parses features and phases with parallel sub-agents, confirms with you before creating anything.
/import-design	Convert a design guide, brand guidelines, or Figma token export into frontend/CLAUDE.md. The ux-review and ux-design agents then enforce your actual design system instead of giving generic feedback.
/dotplanning	Plan a release milestone before development starts. Resolves the scope, asks the twelve kickoff questions that decide the release’s shape (charter, date-vs-scope, capacity, reserve, maturity, freeze, cut line), builds a feature→asset map, flags every missing screen/flow/endpoint/model/doc, sequences the work into gated workstreams, runs a first grooming pass proposing a release:: value for every issue and fitting the committed set to the stated capacity, and drops a self-contained HTML report in ~/Downloads. The begin-gate bookend to /pre-release. Run once per milestone, before the first feature branch.
/adr	Create an Architecture Decision Record in docs/adr/ with sequential numbering. Uses the Nygard template (Context, Decision, Consequences). Good first ADRs: stack choice, data storage, auth approach. Carries two rules that only bite later: Accepted is about the decision, so when it does not mean shipped say so in an Implementation status blockquote (invisible to a status parser by construction, which is the whole point of the shape); and when an ADR moves to Accepted, re-read the open issues that name it — an issue arguing for an ADR is written before it settles, and the delta between the two is exactly the set of rejected options. scripts/adr-accepted-issue-sweep.py is the mechanical half; it is deliberately not a CI gate.

Development commands

Command	What it does
/voc	Voice of the Customer panel — runs each persona through a structured review of a proposed feature. Returns excitement scores, concerns, blockers, and a compact summary to hand off to the architect agent.
/review	Code review against project conventions for a specific file or the full branch diff. Checks correctness, conventions, consistency, testing gaps, and documentation gaps.
/mr	Open a merge request with pre-flight checks: verifies changelog fragment exists, branch naming is correct, no uncommitted changes, no duplicate MR. Generates a structured description with summary, changes, test plan, and notes.
/fix-mr	Get a blocked MR to green and mergeable. Reads both axes — pipeline status and detailed_merge_status — because a green pipeline on a branch that conflicts with the default branch is not ready. Resolves conflicts (checking the worktree for an unpushed resolution first), diagnoses pipeline failures with parallel sub-agents, applies fixes, commits, pushes, and re-checks. Stops after 3 iterations, or when both axes agree. Never merges — only fixes.
/ci-debug	Diagnose a failing pipeline without fixing it. Fetches failed job logs, identifies the stage and root cause, and prescribes a specific fix with the exact file and change needed.
/kaizen	Audit the development harness (CI, gate chain, pre-push, MR flow) for friction and propose a small, ranked list of speed wins. Distinct from /pre-release, which audits the code — kaizen audits the process. Proposes changes and optionally files tooling issues; never edits the harness itself.
/voc-audit	Voice of the Customer panel against a surface that has already shipped, with every finding verified against the code before anything reaches the tracker, and survivors cross-referenced against existing issues. --calibrate scores what the panel predicted against what real users reported. Distinct from /voc, which evaluates a spec before it is built.
/incident-postmortem	Postmortem one resolved harness or process incident: reconstruct the timeline from the mechanical trail, separate root cause from the contributing factors that let it do this much damage, and check whether a gate should have fired. The deliverable is the durable memory entry, not the write-up — a postmortem that lives in a transcript changed nothing. Reactive and incident-scoped, where /kaizen is periodic and pattern-scoped.
/tracker-hygiene	Sweep the tracker for the drift that accumulates between milestone kickoffs: scoped-label exclusivity violations, issues in a dated milestone with no release:: value, likely duplicates, and stale unmilestoned intent. Reports only — it never relabels or closes anything, because the commitment value is the user’s call every time. Cheap enough to run repeatedly, which /dotplanning deliberately is not.
/memory-audit	Sweep the Claude Code memory store for what makes it unrecallable: files with no description: (invisible to recall), files holding several facts, index lines that are bare pointers, dead wiki-links, and grounding claims naming files or flags that no longer exist. Reports first; applies only the safe fixes on confirmation.

Release commands

Command	What it does
/mass-merge	Land a batch of already-green MRs without reddening the default branch. Stacks them on a local integration branch and re-runs the whole-tree aggregate gates after each add (they do not compose across MRs — two changes each under a ratchet ceiling can be over it combined), then merges only the safe prefix serially, bringing each up to the latest default branch by the operation the forge itself will use (MERGE_METHOD). User-invoked only: it force-pushes and merges.
/batch	Land a wave of milestone issues in parallel: one scripts/wt worktree and one delegated agent per issue, each finishing with its own MR. Asks for the milestone and labels rather than inferring them, defaults agents to Sonnet with four stated escalation criteria, caps the wave, and verifies every agent’s commits before believing its “done”. Its rules come from an upstream measurement in which cache reads were 95.9% of subagent spend — the bill is turn count × context size × fan-out. User-invoked only: it spawns agents and spends real money.
/sunset-check	Decide whether an existing surface should be removed, fixed, narrowed, or demoted. Inverts the /voc question — “this is gone next release, what breaks for you?” — and scores removal cost rather than adoption. Reads the code before scoring, because a panel handed a design document scores a feature that may not exist. “Leave it exactly as it is” is not one of its four verbs.
/pre-release	Cross-cutting audit before a release. Runs security, performance, regression, and documentation checks against the full change inventory since the last tag. Reports findings as blocking / important / informational.
/release	Full release flow: version suggestion based on changelog, pre-flight checks, changelog assembly, version bumps, release branch creation. The release goes through an MR like everything else.