Master Conductor: Open‑Source Multi‑Platform AI Agent Skills for Context‑Driven Development

Overview

Conductor is an open‑source, multi‑platform AI‑agent skill derived from Google’s Conductor (a Gemini CLI extension). It supports Claude Code, Gemini CLI, OpenCode, Codex and other CLI tools, and follows a context‑driven development approach: project context is stored persistently in plain Markdown and TOML files rather than temporary chat sessions.

Core workflow

Step 1 – Project setup

Run the setup command at the repository root:

# Claude Code / OpenCode / Codex / Antigravity
set up conductor

# Gemini CLI
/conductor:setup

Conductor asks a series of questions (product vision, target users, main features, tech stack, workflow) and writes the answers to a set of files under conductor/:

conductor/
├── product.md               # Vision, goals, scope
├── product-guidelines.md    # Design and branding rules
├── tech-stack.md            # Stack and tool selection
├── workflow.md              # Development workflow (e.g. TDD)
├── code_styleguides/        # Language‑specific style guides
└── tracks.md                # Index of all tracks

Existing (brownfield) projects are auto‑detected; new (greenfield) projects start from scratch.

Step 2 – Create a new track

When a new feature or bug is ready, create a track:

# Claude Code / OpenCode / Codex / Antigravity
create a new track for dark mode

# Gemini CLI
/conductor:newTrack "Add a dark mode toggle to the settings page"

The command generates two core files inside conductor/tracks/<track_id>/: spec.md – a concrete specification describing what and why. plan.md – a phased implementation plan that breaks the work into phases and tasks.

Conductor leverages any existing context (product definition, tech stack, workflow) to suggest sensible defaults, speeding up the specification process.

Step 3 – Implement the track

After confirming the plan, run:

# Claude Code / OpenCode / Codex / Antigravity
implement the next task

# Gemini CLI
/conductor:implement

Conductor walks through the phases defined in plan.md. For each task it:

Selects the next pending task.

Executes the project’s workflow (e.g. TDD: write failing test → implement → test passes).

Updates the task status back to plan.md.

Prompts for a manual review at the end of each phase.

Commits the code when a task or phase finishes.

All status files ( conductor/tracks.md, conductor/tracks/<track_id>/plan.md, etc.) are persisted, allowing the process to be paused and resumed without losing progress.

Management commands

Status : /conductor:status – summarizes current tracks, phases, and completion.

Revert : /conductor:revert – performs a Git‑aware rollback based on logical units (track, phase, task) rather than raw commit hashes.

Review : /conductor:review – checks completed work against plan.md and product-guidelines.md for compliance.

Installation

git clone https://github.com/jnorthrup/conductor2.git
cd conductor2
# Windows (Git Bash)
sh ./install.sh
# macOS / Linux
chmod +x install.sh
./install.sh

After installation, restart your AI tool to load the Conductor skill.

Practical tutorial – Building a Todo app with Claude Code

Step 1 – Project setup

mkdir todo-app
# Inside the project root, launch Claude Code and run:
设置 conductor
# or
/conductor 设置

Conductor analyses the empty project and asks for product vision, style guide, tech stack, and workflow. The generated files ( product.md, product-guidelines.md, tech-stack.md, workflow.md) become the persistent context.

Step 2 – Create a track

/conductor:newTrack "Add user registration with email verification"

The tool creates spec.md and plan.md for the registration feature, then updates conductor/tracks.md with the new entry.

Step 3 – Implement

/conductor 开始实现

AI follows the TDD flow, committing after each task. Current status can be inspected with: /conductor 查看状态 After each phase, the AI pauses for a manual verification step.

Review and rollback

/conductor 审查我的工作

If the result is unsatisfactory, run: /conductor 还原 Conductor generates a reverse commit that undoes the logical unit without affecting unrelated changes.

Key concepts

Single source of truth – project context lives in version‑controlled Markdown/TOML files.

Track – high‑level work unit containing spec.md and plan.md.

TDD workflow – each phase includes a verification checkpoint.

Git‑aware rollback – revert by track/phase/task, not by raw commit hash.

Token management – large projects consume more AI tokens; a .claudeignore file can exclude irrelevant files, similar to .gitignore.

Best practices & tips

Keep the conductor/ directory under version control; treat it as part of the project.

Always confirm the plan before implementation to avoid drift.

Name tracks clearly; the directory structure mirrors the logical work breakdown.

Define code‑style guidelines early to reduce later refactoring.

Adjust the workflow (e.g., TDD, commit strategy) to fit your team’s needs.

Use .claudeignore to limit token consumption for large codebases.

Common questions

How to resume a track after interruption?

Run implement again; Conductor continues from the last saved state.

Can multiple tracks run in parallel?

It is recommended to work on one track at a time for clear context. Parallel work can be done on separate Git branches.

How does rollback work?

Use the revert command; Conductor creates a Git‑aware reverse commit based on the logical unit (track/phase/task) rather than a raw commit hash.

Should track files be version‑controlled?

Yes – commit the entire conductor/ folder.

Are nested tracks supported?

No. Complex features should be split into multiple phases inside plan.md.

Can the workflow be customized?

Yes – edit conductor/workflow.md to define your own rules.

Reference

GitHub repository:

https://github.com/jnorthrup/conductor2