Skip to content
Skills Library
AI & agents · Planning

parallel-work

/parallel-work

Run several tasks at once, each in its own git worktree with its own Claude Code session, coordinated from the main checkout's session (named coordinator). Use when the human asks to start a task in parallel or in a new worktree/session, to check on or message the other sessions, to reopen one, or to finish a task (PR merged: remove the worktree, fold its notes into AGENTS.md).

Licence
GPL-3.0
Last repository change
11.11.2026

Assessment and security

100/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content35 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
1
Scripts
0
HTML comments
0
URLs in SKILL.md
0
Deprecation

No signals detected. Repository changed on 11.11.2026, without references to retired models or APIs.

Licence

GPL-3.0, otwarta.

What it does and when to use it

Run several tasks at once, each in its own git worktree with its own Claude Code session, coordinated from the main checkout's session (named coordinator). Use when the human asks to start a task in parallel or in a new worktree/session, to check on or message the other sessions, to reopen one, or to finish a task (PR merged: remove the worktree, fold its notes into AGENTS.md).

Original author description

Run several tasks at once, each in its own git worktree with its own Claude Code session, coordinated from the main checkout's session (named coordinator). Use when the human asks to start a task in parallel or in a new worktree/session, to check on or message the other sessions, to reopen one, or to finish a task (PR merged: remove the worktree, fold its notes into AGENTS.md).

How it works

  1. Roles
  2. Start a task
  3. Check on, message, reopen
  4. Finish a task (its PR merged, or dropped)
Original SKILL.md headings · 73 lines, 875 words
SKILL.md preview
---
name: parallel-work
description: >-
  Run several tasks at once, each in its own git worktree with its own Claude Code
  session, coordinated from the main checkout's session (named coordinator). Use when the
  human asks to start a task in parallel or in a new worktree/session, to check on
  or message the other sessions, to reopen one, or to finish a task (PR merged:
  remove the worktree, fold its notes into AGENTS.md).
---

# Parallel work: worktrees and sessions

`tools\debug\worktree.ps1` makes one worktree per task (its own branch, `Play\`, `autotest\`, session) and opens a
Windows Terminal tab running `claude -n <name>` there. What a worktree contains, the shared links and the game-run slots
are described at the top of that script and in `docs/testing.md` section 10. This skill is the coordinator's side.

## Roles

- **coordinator**: the session in the main checkout, named `coordinator` (`worktree.ps1 open -Name main`, or
  `/rename coordinator`). Not `main`: in SendMessage, "main" always means the sender's own conversation, so workers
  could not reach it (not even as `main [ref]`). It stays on `main` and edits nothing in the worktrees: it creates and
  removes them, starts sessions, checks on them, and after a merge updates `main` and folds lasting facts into
  `AGENTS.md`.
- **workers**: one session per worktree, named like the worktree. Each works only in its folder, keeps its notes in
  `TASK.md`, commits and pushes to its own branch without asking, and sends every question, decision or approval
  request (PRs, options, visual checks) and its done/blocked status (with what it committed and pushed) to
  `coordinator`, not to the human in its tab (AGENTS.md: "Questions go through the coordinator").
- The coordinator answers a worker itself when the answer follows from the rules, the task's goal or earlier
  decisions, or is a routine technical choice: "Coordinator's answer: ...". What is the human's call (opening,
  readying, merging or closing PRs, releases, scope or design, player-visible behaviour, deleting things) or needs the
  human's eyes or hands, it asks the human and relays: "Human's answer: ...". Only that counts as the human's
  approval.
- Permission prompts appear in the worker's own tab; when the human types in a worker's tab, the worker answers there.

## Start a task

1. Pick a short name (folder and session name) and a branch: `feat/...`, `fix/...`, `tools/...`. Fix-only PRs start from
   `origin/main` (the default `-From`).
2. Write the goal so a fresh session can start without this conversation: the issue number, what "done" means, decisions
   already made with the human, and limits (fix-only PR, what not to touch).
3. Run it (no `-Build`: the worker does its own first full build, so this returns in seconds):

   ```powershell
   powershell -ExecutionPolicy Bypass -File tools\debug\worktree.ps1 new -Name sliders -Branch fix/f10-sound-sliders -Task "Fix #37: ..."
   ```

   The goal goes into the worktree's `TASK.md` and the new session starts on it. `new` also links the worktree's
   `.claude/skills` to its `.agents/skills` (`tools/setup-agents.py`); `open` does the same, so a worktree that has merged `main` gets its link when it is reopened (or run the script by hand). Quote the goal; keep `;` out of it
   (Windows Terminal splits its command line there).

   Optional: `-Model sonnet|opus|<full model ID>` and `-Effort low|medium|high|xhigh|max` pick the worker's model and
   effort (`claude --model/--effort`); left out, the session uses the defaults. They apply to that session only, nothing
   is remembered.

`worktree.ps1` is a thin wrapper over `tools\debug\worktree.py` (Python 3.8+); the commands and options here are the
same either way.

## Check on, message, reopen

- `worktree.ps1 list`: every worktree with its branch, changed files and who holds a game-run slot. Game runs of the PowerShell scripts and of the Python test kit
  (`tools/testkit/`, `tools/smoke-test.py`) share the same run slots. Not built yet: the full sweep before a release (`.agents/REPO.md`, Testing).
- ListAgents: which sessions are running. SendMessage to `<name>` to ask for a status or pass on a decision from the
  human; the answer arrives as a message. Don't report a worker's state you haven't heard back.
- A closed tab: `worktree.ps1 open -Name <name>` starts a new session that picks the task up from `TASK.md` (so keep
  workers writing their Status there). `-Continue` tries `claude -c` (the folder's last conversation) instead. `-Model`/`-Effort` work here too but are
  not remembered: pass them again on every `open`, or the session starts on the defaults. `new` and `open` start the tab
  without the coordinator's Claude Code variables, so each worker is its own session (in ListAgents, conversation saved).
- The human tests a worktree's build with its own `Play\FrightWorks Game.exe`, or its launcher (the worker runs
  `tools\build-launcher.ps1` there, then `launcher\publish\FrightWorks.exe`). The human's own `Play\` is untouched.
- When one PR merges and another worktree's branch needs it: ask that worker to merge `origin/main` into its branch.

## Finish a task (its PR merged, or dropped)

1. Make sure its session is done: ListAgents, or ask it. Ask the human to close its tab (removing the folder under a
   running session breaks that session).
2. Read its `TASK.md`. Fold what every agent should know into `AGENTS.md`, or into `.agents/REPO.md` when it is a public rule or fact
   (that file changes in the PR that makes it true, not afterwards; case histories go in the skills, via a kit PR); `TASK.md` is deleted with the folder.
3. `worktree.ps1 remove -Name <name>`. It refuses while there are uncommitted or untracked changes, a program running
   from the folder (it stops leftover dotnet build servers itself) or the worktree's Claude session (`-n <name>`). Never delete a worktree folder by hand.
4. Update `main` (`git pull --ff-only`) and delete the merged local branch (`git branch -d <branch>`; GitHub already
   deleted the remote one). A dropped branch: ask the human before `git branch -D`.

Versions

RepositoryScoreLicence
This page · FrightWorks/FrightWorks-no100f100GPL-3.0
DannyKuk/parallel-work100MIT
JQ-WCoding/devkit100MIT
arxdsilva/ai-loops100MIT

Database snapshot: 13.11.2026 · package version 619389aa6b34 (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 100git-worktree
    Use when the user asks to work in a worktree or to isolate the agent's work from other agents in the same checkout.
  • 100session-coordinator
    Coordinate work across separate Claude Code sessions that the user can open, watch and steer - split a task, start one session per piece (each in its own worktree, with a model and effort suited to it), follow their reports, get each branch reviewed, and bring the results…
  • 100dispatch
    Use when a session's job is to hand work out to other agents rather than do it — running a set of tasks across several projects or worktrees, a few at a time.
  • 100multi-session-dev
    用多个 AI 编码会话(Claude Code、claude.ai、Codex 等)并行开发一个或多个仓库时的分工与交接做法:一个管理会话做决策和集成,几个开发会话各在自己的分支和 worktree 里实现、验收、交付,子代理做短调研和审查,云端会话在用户不在时推进。管:什么时候自己做、派子代理还是开新会话;派工前查写入边界、共享契约和版本号、桌面端口数据目录这类只有一份的资源、有几件事在等用户;任务简报;按改动挑验收证据(静态检查、截图或录屏、真机);结束报告第一句和换会话交接;管理会话核对 Git…
  • 90worktree
    Isolate parallel work in a git worktree of the current repository (worktree tool: status / new / merge / remove). Use when another live session is dirty in the same directory (list_agents), when two agents would fight over index.lock, or when starting a parallel task checkout.
  • 100coordinator
    Coordinate one worktree of an hcoord project inside Herdr. Turn the person's requests into small tasks, start one worker agent per task (each on its own branch and checkout, merged back into this worktree when done), react when workers report, and tell the person only what needs…

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.