Skip to content
Back to search

manager-skill

MarketplaceWorks with Claude CodeGitHub

Manager mode for Claude Code: plan, delegate to the cheapest capable model, verify.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/andreacw96--manager-skill/badge.svg)](/listings/andreacw96--manager-skill)

HTML

<a href="/listings/andreacw96--manager-skill"><img src="/api/listings/andreacw96--manager-skill/badge.svg" alt="findsafeskills safety rating" /></a>

Install

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Requires `git`.

In Claude Code:

/plugin marketplace add andreacw96/manager-skill
/plugin install manager@manager-skill

To install it without the plugin system, copy the skill folder by hand:

git clone https://github.com/andreacw96/manager-skill /tmp/manager-skill && mkdir -p ~/.claude/skills && cp -r /tmp/manager-skill/skills/manager ~/.claude/skills/

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json · Manifest · 14 lines
  1. {
  2. "name": "manager-skill",
  3. "description": "Manager mode for Claude Code: plan, delegate to the cheapest capable model, verify.",
  4. "owner": {
  5. "name": "Andrea Tombolato"
  6. },
  7. "plugins": [
  8. {
  9. "name": "manager",
  10. "source": "./",
  11. "description": "Manager mode: Claude plans, briefs, verifies and reports while sessions or subagents do the work."
  12. }
  13. ]
  14. }
README.md · README · 56 lines
  1. <p align="center">
  2. <img src="https://raw.githubusercontent.com/andreacw96/manager-skill/main/docs/brand/banner.png" width="500" alt="manager, a Claude Code plugin" />
  3. </p>
  4. # manager
  5. A Claude Code skill that makes Claude delegate instead of doing the work itself. Claude plans the task, writes a brief for each piece, and hands each one to a separate session or subagent running the cheapest model that can handle it. When a worker reports back, Claude checks the evidence and tells you what landed.
  6. Most of the work runs on haiku and sonnet, which costs less than running everything on your main model. Your main session's context also stays small, and a task only counts as done after Claude has checked it against `git`.
  7. ## Install
  8. Requires `git`.
  9. In Claude Code:
  10. ```
  11. /plugin marketplace add andreacw96/manager-skill
  12. /plugin install manager@manager-skill
  13. ```
  14. To install it without the plugin system, copy the skill folder by hand:
  15. ```bash
  16. git clone https://github.com/andreacw96/manager-skill /tmp/manager-skill && mkdir -p ~/.claude/skills && cp -r /tmp/manager-skill/skills/manager ~/.claude/skills/
  17. ```
  18. ## Use
  19. Say "manager mode" or "delegate everything", or in Italian "modalità manager" or "fai il gestore". As a slash command it is `/manager:manager` when installed as a plugin and `/manager` when copied by hand. Say "stop manager" or "normal mode" ("basta gestore", "modalità normale") to turn it off.
  20. Then describe the work as usual, for example: *"Bump the SDK in api and web, fix the breaking changes."*
  21. ## How it works
  22. | Step | What Claude does |
  23. |------|------------------|
  24. | Plan | Splits the request into independent pieces, one per repo or area |
  25. | Brief | Sends each worker the same template: goal, path, constraints, a concrete check for "done", where to stop, what to report |
  26. | Route | Picks the model by how clear the task is. A large but fully specified task still goes to a cheap model: haiku for mechanical work, sonnet by default, opus for unclear or risky work, fable to review opus-tier changes before merge |
  27. | Verify | Treats each report as a claim and checks the branch, commit, diff stat and tests before calling it done |
  28. | Escalate | If the brief caused a failure, fixes it and retries once on the same model. Otherwise moves haiku to sonnet after one failure, and sonnet to opus after two |
  29. Workers can't run on a higher model than your main session, so start it on opus if you want opus available.
  30. Claude doesn't edit project repos itself unless you say "do it here" for a specific task. It runs at most 4 workers at a time, with one writer per repo (two only if they touch different files). Workers stop at a commit on a local branch. Nothing is pushed, opened as a PR or merged until you approve.
  31. ## Compatibility
  32. In the Claude desktop app (Code tab), each implementation task runs in its own session and worktree, and you can open and steer any of them. Small edits and searches go to subagents.
  33. In the CLI, IDE extensions and other clients, all workers are subagents (the `Agent` tool with worktree isolation). The rules and briefs are the same, but you can't open the workers as separate sessions.
  34. ## License
  35. MIT

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.