Opus Manager skill
A skill that makes Claude a manager: it writes tickets, hands code to cheaper agents, verifies and asks another vendor to review.
- 63 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The skill runs workers unattended in your repository and spends quotas of several services.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/opus-managerDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The skill runs workers unattended in your repository and spends quotas of several services.
This is third-party code. Review the repository files before installing.
What it does
The `opus-manager` skill for Claude Code changes Claude's role: instead of typing code it plans work as tickets, hands them to cheaper CLI agents on the machine, verifies the results and keeps the queue going while the human is away. A ticket with a goal, an allowlist of files and acceptance commands lives in `_tickets/open/`, dispatch moves it to `doing` and the worker leaves a receipt in `_receipts/` with the exact commands and their output. Claude reruns the acceptance commands itself: a receipt is a claim, not evidence. Then a model from another vendor reviews the change read-only and reports findings, Claude checks each one and real ones go back as a fix ticket. Everything is Markdown in the repository. Decisions it cannot make alone pile up in a list for the human and a short report with costs comes at the end.
Who it is for. Claude Code developers on a limited plan who want to leave tasks running unattended and spend their quota on planning and acceptance.
Good fit when
- A queue of tickets is set in the evening and a report with results arrives in the morning.
- Implementation goes to a cheap model while Claude keeps acceptance and review.
- A change should also be checked by a model from another vendor.
- Claude context should not fill with code, test logs and failed attempts.
Not a fit when
- No other CLI agents are installed on the machine to hand code to.
- Workers must not run commands without your supervision.
- The task is small and easier to fix directly.
Example request
Build a ticket queue for these three tasks, hand them to the workers, verify the results yourself and send a short report with costs.Limitations
Work stops if the computer sleeps, a permission prompt goes unanswered or Claude hits its limits: on the Pro plan one busy morning used about 83% of a five-hour window. Workers already started finish and you can continue after the reset. Workers run as separate unattended processes in your repository, so file limits and acceptance commands in tickets must be strict. A model from another vendor is needed for review. The savings examples come from the author's descriptions and were not measured independently.
How to disable. Delete the `~/.claude/skills/opus-manager` folder and restart Claude Code. Clean the `_tickets` and `_receipts` directories in projects separately.
FAQ
How do I install it?
Ask Claude Code to install `skill/en/opus-manager` from the repository as a personal skill or copy the folder into `~/.claude/skills/` by hand.
Which agents take the code?
Cheaper CLI agents already on the machine, such as DeepSeek and GLM in the author's examples.
Why is a receipt not trusted?
A worker only claims a result, so Claude reruns the acceptance commands and reads the diff itself.
Is there a Chinese version?
Yes, the skill also lives in `skill/zh/opus-manager`.
What can stop the work?
The computer going to sleep, an unanswered permission prompt and exhausted Claude limits.
Discussion
No questions or reviews yet.
Related
SkillsClaude Code supervises while Codex, opencode and pi work on your quota: delegated edits, independent review, model discussions
31 skills for Claude Code and Codex: cross-vendor review, task handoff on low quota, a morning brief and a nightly retrospective
A Claude Code skill that hands implementation-heavy work to Codex CLI while Claude keeps planning, review and acceptance.
A personal ~/.claude for Claude Code: twelve subagents, review commands, worktree and goal skills and a Go status line.
A skill that turns the main Claude Code session into a dispatcher: tasks go to background subagents and the chat never blocks.
A skill that makes the lead Claude model a foreman: it plans work, hands tasks to worker agents and verifies the result itself.
README in short
The README covers the manager-instead-of-typist idea, two real runs, cost reasoning, steps from ticket to finding verification, installation and first run.