claude-kit

Work a ticket as a build-broker worker — edit in a cheap git worktree (never build), commit with explicit paths, submit a job to the shared broker, wait for the result, and land through it. Use when many agents share one build checkout via the broker (scripts/broker/) instead of each spinning up its own environment.

Capture and route an interjection into the active project's .ai store without derailing current work. Use when the user says to log, capture, remember, defer, or file an idea, bug, question, decision, or note.

Manage work and tooling through the claude-kit repo — the canonical, cross-machine source of truth for the Claude Code and Codex workflow (.ai/ capture→triage→work→drain→flush), enforcement hooks, and shared commands, agents, and skills. Use when setting up a project, driving the backlog, or maintaining reusable workflow tooling.

Review a diff or recent code changes for correctness, security, maintainability, and missing tests without editing. Use after implementation or when the user asks for a code, diff, branch, or pull-request review.

The DB & ORM checklist — mandatory before writing ANY database code (SQL, schema/migrations, Drizzle, Alembic, ORM queries). Load it the moment DB work appears; the hooks enforce the top violations, this skill carries the full per-engine defaults.

Batch unresolved project decisions into a structured questionnaire and persist the answers. Use when the user asks to decide open questions, resolve pending choices, or run the decision queue.

Audit a repo's docs against the actual code — broken links, stale references to renamed/deleted files or symbols, undocumented public surface, and out-of-date examples. Use after a refactor/rename, before a release, or when docs feel untrustworthy. Reports findings with pointers; you approve fixes.

Close a reviewed claude-kit ticket with explicit human acceptance. Use only when the user asks to mark a ticket done, accepts UAT, or approves closure of a review-stage item.

Select and execute the next ready claude-kit backlog item according to .ai/config.yml. Use when the user asks to drain the queue, keep going, pull the next task, or work the backlog without naming a ticket.

Persist the active task, exact resume state, and settled decisions before compaction, clearing, or ending a session. Use when the user asks to flush, checkpoint, save context, compact safely, or prepare a handoff.

Build procedural game assets in code from shaped primitives and reusable component recipes, with low-poly vertex-level modelling, shader-first surfaces (bake only what the shader cannot do, owner and wear as live instance parameters, one surface per design, meshes shared by shape) and numeric validation. Use for props, vehicles, buildings, characters, terrain assets, or kitbash libraries.

Spin up a greenfield Rust game POC on the rapid-game framework — one command to a running Bevy app with the editor shell, hot-reload tunables, headless probe, and a wasm-clean web path. Use when the maintainer wants a new game POC or prototype started.

Reconstruct a human-readable briefing from durable claude-kit state across one or more projects. Use when starting or resuming work, catching up after a gap, or asking what currently needs attention.

Perform a focused behavior-preserving refactor and prove equivalence with tests. Use for extract, rename, split, duplication removal, or layer untangling when behavior must remain unchanged.

Run a disciplined pre-push / pre-release verification before publishing a change — tests, build, debug-leftover scan, version/changelog, and commit hygiene. Use before opening a PR, pushing to main, or cutting a release. Reports a go/no-go with the evidence; fixes nothing on its own.

Conduct a read-only investigation across a codebase, git history, or authoritative web sources and return a synthesized answer with pointers. Use when a question requires tracing many files or gathering sourced external evidence.

Scaffold a committed project-local knowledge agent for Claude Code, Codex, or both, pre-filled from verified domain conventions and past gotchas. Use when a domain keeps recurring across delegations and needs a durable agent under .claude/agents or .codex/agents.

Show a read-only cross-project claude-kit status briefing without resuming or mutating work. Use when the user asks for a standup, status glance, current queue summary, or what needs their attention.

Add or extend runnable automated tests for a behavior, bug, or untested module and execute them. Use when the user asks to add tests, reproduce a regression, lock in a fix, or establish test-backed evidence.

Drain claude-kit inbox captures into tickets, questions, decisions, and notes, then report the prioritized worklist. Use when the user asks to triage captures, process the inbox, or organize incoming work.

Pick up one named claude-kit ticket, restate its acceptance criteria, confirm scope, implement it, and stop at review. Use when the user asks to work, implement, or resume a specific ticket ID.