| name | parallel-dispatch |
| description | Plan and split work into parallel worktree sessions. Use whenever the user wants to work on multiple tasks at once, mentions parallel sessions, worktrees, "split this up", "dispatch", or lists several independent tasks for one work session. |
parallel-dispatch
Plan-then-spawn discipline: turn a task list into at most 3 scoped worktree sessions, each launched by the human with a tight dispatch brief. Mechanics of worktrees themselves (what carries over, cleanup, hard rules) live in docs/parallel.md — read it before dispatching; do not restate it here or to the user.
Purpose
Parallel sessions are only faster when they never touch the same files and never have to rediscover their scope. This skill front-loads that thinking into one dispatcher session, then hands each worktree a brief so it starts working instead of exploring — that's the token-efficiency core.
Process
1. Collect the task list
From the user's message, docs/STATE.md "Now" section, or Linear (if connected). Confirm the list with the user before splitting.
2. Dependency check — the gate
For each pair of tasks, identify which files/modules each will touch (search now, in this session). Tasks with overlapping file surface MUST be serialized into the same session, never parallelized — overlapping worktrees produce merge conflicts, not speed. When unsure whether surfaces overlap, treat them as overlapping.
3. Cap parallelism
At most 3 parallel sessions (4 absolute max, only with explicit user push). More cannot be supervised, and unsupervised output is not high-end output. Extra tasks queue for the next round.
4. Emit one DISPATCH BRIEF per parallel task
The user pastes it as the first message in that worktree session. Max 12 lines:
# Dispatch brief — [kebab-task-name]
Task: [one sentence]
Scope: [exact files/directories in play]
Do not touch: [every file another parallel session owns, plus docs/STATE.md]
Definition of done: [tests/checks that must pass]; deliverable = a PR
Linear: [issue ID, or "none"]
Your scope is this brief plus what loads at startup; don't explore beyond it.
5. Print launch commands
One per task, with its brief:
claude -w <kebab-task-name>
Remind the user to spot-check /context in the first worktree session for Plugin (ravie) entries — the missing-skills failure mode and its fix are in docs/parallel.md.
6. Dispatcher responsibilities
This session (main checkout) owns docs/STATE.md updates and final merges. Merge the returning PRs sequentially, re-running checks after each merge — never batch-merge.
Output format
Task table (task → session assignment → file surface), then the briefs, then the launch commands. Nothing else.
Hard rules
- NEVER spawn sessions, run
claude -w, or create worktrees yourself — this skill plans and emits briefs; the human launches terminals
- Never parallelize tasks with overlapping file surface; when unsure, serialize
- Never exceed 4 sessions; default cap is 3
- Every brief's "do not touch" list must include every file another parallel session owns
- STATE.md is edited only from this dispatcher session (see
docs/parallel.md hard rules)
Connects to
docs/parallel.md — worktree mechanics, plugin-scope failure mode, cleanup, merge-first dependencyissue-to-pr — each worktree session typically runs it against its briefgithub-operator — sequential merges of the returning PRs
Common failure modes
Optimistic splitting — two tasks "probably" touch different files; nobody checked; both edit the same component and the merge burns the saved time. Do the dependency check with actual searches.
Scope-free sessions — a worktree session launched with just "fix the dashboard" spends its first 20k tokens rediscovering context the dispatcher already had. Always paste the brief.
Fan-out pride — six parallel sessions, zero supervision, six mediocre PRs. Three supervised beats six abandoned.
