A skills plugin that organizes AI-assisted software development around Fred Brooks' Surgical Team model from The Mythical Man-Month (1975). Compatible with GitHub Copilot CLI, Claude Code, and OpenCode.
Instead of every team member working on all parts of a system, the Surgical Team concentrates critical work in one skilled "surgeon" (chief programmer), supported by specialized roles that keep the surgeon focused and productive. This plugin maps those roles to agent skills and dispatch templates.
This project draws from two sources:
Fred Brooks' Surgical Team (Chapter 3, The Mythical Man-Month, 1975), originally conceived by Harlan Mills: a small, highly specialized team organized around a single chief programmer who writes all critical code, supported by a copilot, tester, toolsmith, editor, administrator, language lawyer, and program clerk — each with a distinct, non-overlapping responsibility.
Superpowers by Jesse Vincent: a Claude Code skills framework that demonstrated how composable, role-aware skills can guide an AI agent through disciplined software development workflows. The structure, conventions, and plugin format of this project follow Superpowers' design closely.
The SKILL.md format used here conforms to the Agent Skills open standard, which is supported by GitHub Copilot CLI, Claude Code, and OpenCode.
| Role | Skill | Responsibility |
|---|---|---|
| Surgeon | surgeon |
Chief programmer — owns all critical decisions and implementation |
| Copilot | copilot |
Alter ego — reviews all of the Surgeon's work against design intent |
| Tester | tester |
Adversarial quality — finds failure modes the Surgeon was too optimistic to anticipate |
| Administrator | administrator |
Project tracking — manages tasks, blockers, and scope so the Surgeon doesn't have to |
| Editor | editor |
Documentation — ensures everything written about the code is accurate and complete |
| Toolsmith | toolsmith |
Automation — builds utilities that make the Surgeon permanently faster |
| Language Lawyer | language-lawyer |
Edge cases — answers the language/framework questions where being wrong causes subtle bugs |
| Program Clerk | program-clerk |
Code organization — keeps the codebase navigable, names accurate, dependencies clean |
You are always the Surgeon. The other roles are invoked as needed — either as inline guidance (the skill guides Claude to perform that role temporarily) or as dispatched subagents (for independent, parallelizable work like code review or test writing).
Register the marketplace, then install the plugin:
copilot plugin marketplace add zakame/skills-marketplace
copilot plugin install brooks-agent-team@zakame-skills-marketplace
Skills are available immediately. To update to a newer version:
copilot plugin update brooks-agent-team
To load skills directly from a local clone — useful when developing or testing changes to the skills themselves:
git clone https://github.com/zakame/brooks-agent-team /path/to/brooks-agent-teamRegister the skills directory in a Copilot CLI session:
/skills add /path/to/brooks-agent-team/skills
The copilot, tester, and language-lawyer custom agents are bundled in .github/agents/. For marketplace installs, they are automatically available via /agents after copilot plugin install — no additional setup needed.
For manual installs, if the agents are not discovered automatically, copy them to ~/.copilot/agents/:
cp /path/to/brooks-agent-team/.github/agents/*.agent.md ~/.copilot/agents/See Dispatch subagent roles for invocation examples.
OpenCode discovers skills from .opencode/skills/, .claude/skills/, and ~/.config/opencode/skills/ (global). Clone this repository to a stable location of your choice, then symlink each skill subdirectory into a discovery path:
git clone https://github.com/zakame/brooks-agent-team /path/to/brooks-agent-team
# Project-level (one-time per project):
mkdir -p .opencode/skills
ln -sf /path/to/brooks-agent-team/skills/* .opencode/skills/
# OR global (available in all projects):
mkdir -p ~/.config/opencode/skills
ln -sf /path/to/brooks-agent-team/skills/* ~/.config/opencode/skills/To update, run git pull inside /path/to/brooks-agent-team.
OpenCode auto-discovers agent definitions from .opencode/agents/ in your project directory. To make the Copilot, Tester, and Language Lawyer subagents available:
mkdir -p .opencode/agents
cp /path/to/brooks-agent-team/.opencode/agents/*.md .opencode/agents/For cross-project use, global agent discovery paths are not yet standardized in OpenCode — check the OpenCode agent documentation for current options.
See Dispatch subagent roles for invocation examples.
Register the marketplace, then install the plugin:
/plugin marketplace add zakame/skills-marketplace
/plugin install brooks-agent-team@zakame-skills-marketplace
Skills are available immediately. To update to a newer version, run /plugin install brooks-agent-team@zakame-skills-marketplace.
To load skills directly from a local clone — useful when developing or testing changes to the skills themselves:
git clone https://github.com/zakame/brooks-agent-team /path/to/brooks-agent-team
claude --plugin-dir /path/to/brooks-agent-teamTo reload after making changes without restarting:
/reload-plugins
| Skill / Command | Tool | When to use |
|---|---|---|
assemble-team skill |
Copilot CLI, Claude Code & OpenCode | Single-session work — one AI instance plays all roles sequentially |
/assemble-team command |
Claude Code only | Same as above, as a slash command |
assemble-with-fleet skill |
Copilot CLI & OpenCode | Parallel work — spawns one independent session per role (Copilot CLI uses /fleet; OpenCode uses the task tool) |
/assemble-with-agent-teams command |
Claude Code only | Parallel work — spawns via Claude Code Agent Teams |
Run at the start of any development session to get a project-contextual briefing on which roles apply and how to invoke them.
Copilot CLI:
Use the assemble-team skill
Claude Code:
/assemble-team
OpenCode:
Use the assemble-team skill
The AI surveys your project and presents a tailored overview of the team. Roles are invoked on demand as the work requires them. Lightweight and works without any additional setup.
Spawn one independent AI session per role so that Copilot reviews and Tester writes tests while you continue on the critical path.
Copilot CLI — uses assemble-with-fleet skill (requires experimental fleet mode):
Use the assemble-with-fleet skill
OpenCode — uses assemble-with-fleet skill (spawns subagents via the task tool):
Use the assemble-with-fleet skill
Claude Code — uses /assemble-with-agent-teams (requires Claude Code Agent Teams):
Enable Agent Teams first:
{
"env": {
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
}
}Then run:
/assemble-with-agent-teams
What you get:
- One independent session per role (Copilot, Tester, and any others you choose)
- True parallelism — Copilot reviews while you continue coding
- Shared task list with dependency tracking and role-tagged tasks (
[implement],[review],[test], etc.) - File ownership assignments to prevent conflicts
Copilot CLI — use the skill name in your prompt:
Use the surgeon skill to start implementing this feature.
Use the copilot skill to review my changes.
Use the language-lawyer skill for this edge case.
Claude Code — skills are namespaced under brooks-agent-team::
brooks-agent-team:surgeon
brooks-agent-team:copilot
brooks-agent-team:tester
brooks-agent-team:administrator
brooks-agent-team:editor
brooks-agent-team:toolsmith
brooks-agent-team:language-lawyer
brooks-agent-team:program-clerk
OpenCode — use the skill name directly in a session prompt:
Use the surgeon skill to start implementing this feature.
Use the copilot skill to review my changes.
Use the language-lawyer skill for this edge case.
The AI reads each skill's description field and automatically invokes the relevant skill when your request matches its trigger conditions. For example:
- Starting an implementation task →
surgeonskill loads - Completing a feature →
copilotskill is suggested - Spotting repeated manual operations →
toolsmithskill activates - Encountering a language edge case →
language-lawyerskill is invoked
The Copilot, Tester, and Language Lawyer roles can be dispatched as independent subagents, allowing the Surgeon to continue working on the critical path while review, test writing, or edge-case research happens in parallel.
Copilot CLI — use the custom agents in .github/agents/ (copy to ~/.copilot/agents/ for cross-project use):
Use the copilot agent to review the authentication changes.
Use the tester agent to write tests for the payment module.
Use the language-lawyer agent to research this API deprecation.
Claude Code — use the dispatch templates in agents/copilot.md, agents/tester.md, and agents/language-lawyer.md.
OpenCode — use the agent definitions in .opencode/agents/:
Dispatch the Copilot agent to review the authentication changes.
Dispatch the Tester agent to write tests for the payment module.
Dispatch the Language Lawyer agent to research this API deprecation.
The Copilot agent is read-only (permissions deny edit, bash, and webfetch); the Tester agent can write files and run shell commands (permissions allow edit and bash); the Language Lawyer can search the web, fetch URLs, and run shell commands (permissions allow bash, webfetch, and websearch, deny edit).
skills/ Shared across all platforms (Agent Skills standard)
using-brooks-team/SKILL.md Meta-skill: team orientation and role routing
surgeon/SKILL.md Chief programmer operating guide
copilot/SKILL.md Code review and backup
tester/SKILL.md Adversarial test strategy
administrator/SKILL.md Task tracking and scope defense
editor/SKILL.md Documentation and accuracy
toolsmith/SKILL.md Custom tool builder
language-lawyer/SKILL.md Language and framework edge cases
program-clerk/SKILL.md Code organization and structure
assemble-team/SKILL.md Team briefing skill (Copilot CLI and OpenCode)
assemble-with-fleet/SKILL.md Parallel team spawn via Copilot CLI fleet mode or OpenCode task tool
.claude-plugin/ Claude Code specific
plugin.json Plugin manifest (name, version, author)
agents/ Claude Code dispatch templates
copilot.md Subagent template for Copilot role
tester.md Subagent template for Tester role
language-lawyer.md Subagent template for Language Lawyer role
commands/ Claude Code slash commands
assemble-team.md /assemble-team command
assemble-with-agent-teams.md /assemble-with-agent-teams command
.github/agents/ GitHub Copilot CLI custom agents
copilot.agent.md Code reviewer agent
tester.agent.md Adversarial tester agent
language-lawyer.agent.md Language and framework edge-case agent
.opencode/agents/ OpenCode custom agents
copilot.md Code reviewer agent
tester.md Adversarial tester agent
language-lawyer.md Language and framework edge-case agent
The skills/ directory is the core of the plugin. Each subdirectory contains a SKILL.md file conforming to the Agent Skills open standard, which is supported by GitHub Copilot CLI, Claude Code, and OpenCode. These files work on any platform that reads the standard.
Platform-specific files provide deeper integration:
| Directory | Platform | Purpose |
|---|---|---|
.claude-plugin/ |
Claude Code & Copilot CLI | Plugin manifest (marketplace loading) |
agents/ |
Claude Code | Dispatch templates for subagent roles (Copilot, Tester) |
commands/ |
Claude Code | Slash commands (/assemble-team, /assemble-with-agent-teams) |
.github/agents/ |
Copilot CLI | Custom agent definitions for /agent dispatch |
.opencode/agents/ |
OpenCode | Custom agent definitions for subagent dispatch |
skills/assemble-team/ |
Copilot CLI & OpenCode | Team briefing skill |
skills/assemble-with-fleet/ |
Copilot CLI & OpenCode | Parallel spawn via fleet mode or task tool |
Note: Copilot CLI recognizes
.claude-plugin/in addition to.github/plugin/when loading plugin manifests.
The eight role skills in skills/ are compatible with OpenCode via the Agent Skills open standard — OpenCode reads SKILL.md files the same way Claude Code and Copilot CLI do. See OpenCode skills documentation for the full discovery path list and SKILL.md format reference.
Full agent dispatch is supported through .opencode/agents/, which includes Copilot, Tester, and Language Lawyer subagent definitions. These files use OpenCode's agent frontmatter format (description, mode: subagent, permission block). Permission keys map to specific tools: edit controls write/edit/patch operations, bash controls shell commands, webfetch controls web fetches, and read controls file reads (allowed by default). The Copilot sets edit: deny, bash: deny, and webfetch: deny; the Tester sets edit: allow and bash: allow. To restrict file reads: read: deny.
If OpenCode updates its agent frontmatter format, check the OpenCode agent specification to verify these files remain current.
Note for maintainers: The agent body content is duplicated across three locations:
.opencode/agents/,agents/(Claude Code), and.github/agents/(Copilot CLI). Any change to the review, test, or language research protocol must be applied in all three.
- One surgeon, many supporters. Productivity comes from keeping the chief programmer focused, not from adding more equal contributors.
- Roles are process gates, not suggestions. When a role applies, invoking its skill is mandatory — it protects the quality of the Surgeon's work.
- Delegate early, not late. Recognizing the need for a supporting role at the start of a task is cheaper than discovering it after struggling.
- Code belongs to the system. Write as if a future Surgeon will read it cold.
MIT License — see LICENSE for details.