Google releases Conductor, a Gemini CLI context-driven extension that saves knowledge in Markdown format and orchestrates workflows.

Google’s Conductor is an open source extension that extends Gemini CLI. This allows AI to be generated in a more structured and context-driven workflow. Conductor saves product knowledge, work plans, and technical decisions as Markdown files in the repository. Gemini agents then use these files to drive their workflows instead of using ad-hoc chat requests.

Context driven development: a shift from chat-based code to contextualized development

The majority of AI programming today is session-based. The context is lost when you end the chat session. You can paste the code in a chat and describe the task. This is a major problem for Conductor.

Conductor uses a persistent directory to store context information instead of the ephemeral prompts. As Markdown, it captures the product goals and constraints as well as workflow rules, style guides, technology stacks, etc. Gemini reads the files every time it runs. The AI behaviour is then repeatable between machines, shells and team members.

A simple life cycle is also enforced by Conductor:

Context → Spec and Plan → Implement

It does not go directly from the natural language to edit code. First, it creates a track and then generates an outline, a plan.

Installing Conductor into Gemini CLI

Gemini’s CLI runs Conductor. Installing Conductor is as simple as:

gemini extensions install https://github.com/gemini-cli-extensions/conductor --auto-update

It is important to note that the word “you” means “you”. --auto-update Flag is optional, and it keeps extension updated with latest version. Gemini CLI will display Conductor commands when in the project directory.

Setup a project with /conductor:setup

Start with the project setup.

The interactive session builds the context. The conductor will ask about users, product, technology stack, development practices, and requirements. These answers are used to generate a conductor/ A directory containing several files For example:

• conductor/product.md
• conductor/product-guidelines.md
• conductor/tech-stack.md
• conductor/workflow.md
• conductor/code_styleguides/
• conductor/tracks.md

They describe the target users, high level features, accepted technologies and testing expectations. The artifacts define the project’s target audience, top-level features, acceptable technologies, expected testing and coding practices. The context files are stored in Git along with all the other source code. This allows for easy review and auditing of changes.

Tracks as First Class Artifacts

Introduces the conductor Tracks The term “features” is used to describe units of work, such as new features or bug fixes. You can create a sound track using:

Or with a brief description

/conductor:newTrack "Add dark mode toggle to settings page"

Each new track is assigned a folder under the Conductor directory. conductor/tracks// containing:

• spec.md
• plan.md
• metadata.json

spec.md This document contains the requirements and restrictions for each track. plan.md This document contains an execution plan that is broken down into tasks and phases. metadata.json Stores identifiers, status and information.

Conductor assists in drafting spec and plan by using existing context files. They are then edited and approved by the developer. All implementations must be based on a clearly defined and version-controlled plan.

Implementation of /conductor:implement

When the plan has been prepared, The agent will take control of the vehicle:

Readers plan.mdSelects the next unfinished task and executes the workflow. The following are typical cycles:

1. Review relevant files in context.
2. Changes to the code are proposed.
3. Check or test according to conductor/workflow.md.
4. Update the task status plan.md Globalization tracks.md.

This extension includes checkpoints placed at the phases boundaries. Conductor will pause at each of these phases to wait for human confirmation before moving on. The agent is prevented from making large, unreviewed changes.

The flow of information is supported by a number of operational commands.

• /conductor:status The task is shown in progress.
• /conductor:review Validates completed work in relation to product and style guidelines.
• /conductor:revert Git allows you to roll back an entire track, task or phase.

In a workflow with multiple changes, reverts are described in terms of track hashes and not the raw commit.

Brownfield projects, teamwork and workflow

The Conductor tool is not limited to new code. You can run Conductor by running /conductor:setup In an existing repository the context session can be used to transform implicit knowledge of the team in Markdown. As more tracks are run over time, the directory will become a compact description of the architecture and constraints within the system.

In the code, team behavior is encoded workflow.md, tech-stack.mdStyle guide and. All engineers or AI agents who use Conductor on that repo will inherit the rules. This can be useful in enforcing test strategy, setting expectations or approving frameworks among contributors.

The context and plan files can be changed in the same way as source code because they are stored within Git.

What you need to know

• Gemini’s CLI Extension Conductor allows context-driven programmingThis is an Apache 2.0-licensed open source extension which runs within Gemini CLI. It drives AI agents using repository local Markdown context, instead of prompts.
• The project context is saved as Markdown versioned under conductor/: Files like product.md, tech-stack.md, workflow.mdThe agent must read the code style guide for each job. It defines product goals, technology choices and workflow rules.
• The work is divided into different tracks spec.md The following are some examples of how to get started: plan.md: /conductor:newTrack Creates a directory of tracks spec.md, plan.mdThen, metadata.jsonMaking requirements and implementation plans explicit and reviewable.
• Controlled implementation of the program is done via /conductor:implement The track-aware operationThe agent performs the tasks as per plan.mdUpdates on progress tracks.mdThe. and Supports /conductor:status, /conductor:reviewThen, /conductor:revert For progress inspection and Git backed rollback.

Click here to find out more Repo The following are some examples of how to get started: Technical details. Also, feel free to follow us on Twitter Join our Facebook group! 100k+ ML SubReddit Subscribe now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.