.claude

Run a body of work as an orchestrator: plan -> partition into file-disjoint waves -> delegate to N parallel code-implementer subagents -> integrate, verify, review -> (PR on request). The main session orchestrates and NEVER writes implementation code itself. Every wave and run is logged for later analysis. Use for non-trivial multi-file features run inside a dedicated worktree.

Ship a completed /orchestrate milestone: open a PR for the worktree branch, poll CI until conclusive (never cancel — self-hosted runners can be slow), auto-merge when every check is green and the PR is MERGEABLE/CLEAN, then clean up the worktree + branches. In project-kit repos the milestone's status.json now->shipped bump rides INSIDE the milestone PR (no direct push to main), so parallel sessions never race a shared-doc write. Run AFTER a successful /orchestrate when the branch's work is committed and verified. It ships what's there — it never writes or fixes code; a red CI or required review ends the skill with a report, not a patch.

Generate a .worktreeinclude file for the current git repository so Claude Code copies gitignored local files (env, secrets, certs) into new worktrees.

Bootstrap a project's agent-facing context system — structured JSON docs (live status + decision records) rendered by a live HTML dashboard, plus the operating rules that keep them current, wired into CLAUDE.md. Use when setting up context docs for a new or existing project, "set up project docs", "scaffold status/ADRs", or first-time project setup. Usage: /project-kit [--minimal | --full] [path]

This skill should be used when the user asks to "wrap up", "wrap session", "finish up", "wrap this work", "close out this work", "I'm done", "finalize this", "let's finish", or "end this work". It orchestrates session finalization: polish loop, commit, push, and mark work done.

Open a markdown file in a browser and leave block-level comments that are saved to a sidecar JSON file alongside it. The skill launches the server in the harness's background (so python is tied to your Claude session, not orphaned) and auto-opens the URL in your default browser. To apply comments after clicking Done, ask Claude to "apply" — if "Auto-apply" was checked in the Submit modal, Claude proceeds without asking for confirmation.

Drafts a technical implementation plan that bridges a PRD to code (approach, file changes, types, tagged task list with per-task tests), runs one optional critic pass via devils-advocate, batches user decisions in one AskUserQuestion. Invoked only via the explicit `/create-plan` slash command.

This skill should be used when the user asks to "execute the plan", "implement the plan", "run the plan", "build from plan", or wants to sequentially implement all tasks from a plan document. Parses plan §4, invokes /execute-task for each task in dependency order, runs stage-end test+validation, then polish (1× or 2× per change-size heuristic).

This skill should be used when executing a single implementation task from a plan document. Runs the task through implement → validate, with retry + model escalation on validation failure. Called by /execute-plan for each task, or directly for re-running a single task.

Drafts a PRD for a feature, runs one critic pass via the devils-advocate subagent, then lets the user direct revisions in one batched interaction. Invoked only via the explicit `/create-prd` slash command.

This skill should be used when the user asks to "polish code", "clean up code", "simplify and review", "refine code", or wants to run code simplification followed by gated code review on recently modified code.

Ask user anwer about claude code