ai-driven-engineering

AI-Driven Engineering Workflow

Standardize and automate your engineering process with Gemini CLI.

This repository houses the Workflow Extension for Gemini CLI—a powerful set of autonomous tools designed to guide you through the entire software development lifecycle, from ambiguous idea to merged Pull Request.

It integrates seamlessly with Linear (for project management) and GitHub (for version control) to keep your focus on shipping value, not managing tickets.

---

100%

Of my code is written by AI.

I run AI-driven engineering workshops with companies worldwide—both remote and on-site. Some invite their whole engineering department, others bring non-technical teams along for the ride.

"Last year summer, I asked myself a question: could I be an engineer without writing a single line of code? Not vibe coding—proper AI Driven Engineering.

Today, my engineering life has been changed completely. I use AI agents for almost everything, and I write 0% of the code myself.

We are moving from a world of asking, 'What code do I write?' to a world of asking, 'What problem do I need to solve?'

You are no longer limited by the syntax you know, the documentation you read, or the libraries you've memorized. You'll be limited only by your imagination and your token budget."

---

📦 Installation

For Gemini CLI:

gemini extensions install https://github.com/SaschaHeyer/ai-driven-engineering --auto-update

For Claude Code: Open Claude Code in your terminal and run:

/plugin install ai-driven-engineering@SaschaHeyer/ai-driven-engineering

Local Development

If you have cloned this repository locally and want to test changes:

For Gemini CLI:

gemini extensions link ./workflow-extension

For Claude Code: You can start Claude Code and point it to your local directory:

claude --plugin-dir /path/to/ai-driven-engineering

Or, add it directly via the CLI:

claude plugin add /path/to/ai-driven-engineering

---

🎯 Motivation

"The entire purpose of this structured workflow is to do the hard clarification and planning work upfront."

This isn't for quick, one-line fixes. This workflow is designed for shipping significant new features or handling complex migrations in large, existing codebases—work that involves multiple files, new logic, and proper engineering effort.

Why? By investing time in the first two steps (Define & Plan), we enable our AI agents to run autonomously for hours with minimal supervision.

We break the workflow into distinct phases: Issue, Task, Implement, and Walkthrough. This structure provides the context and boundaries the agent needs to execute complex work without constantly asking, "What next?"

---

🌳 The Secret Weapon: Git Worktrees

"You can't have agents working in the same folder; they would just overwrite each other's work."

To enable true parallel autonomy, we leverage Git Worktrees. This feature allows you to check out multiple branches from a single repository into separate directories.

Imagine one repository, but with 10, 15, or 20 different features and bugs, each living in its own clean, isolated folder.

• Agent A works on feature-1 in worktrees/feature-1.
• Agent B fixes bug-2 in worktrees/bug-2.

They run in parallel, on the same codebase, completely isolated. No conflicts. No stash hell. You can test each agent's work in its own dedicated folder.

This workflow is optimized around standard git worktree commands, which are handled automatically for you by the included git-worktree skill.

The /implement command linked here is designed to handle this isolation automatically.

---

🚀 The Workflow

We follow a strict Define → Plan → Build → Walkthrough → Ship cycle. This extension provides a specialized AI agent command for each stage.

graph LR
    A[Start] --> B(/issue);
    B --> C(/tasks);
    C --> D(/implement);
    D --> E(/walkthrough);
    E --> F(/finalize);
    F --> G[Merged];

Stage	Command	Description
1. Define	/issue	Turns a rough idea into a comprehensive Product Requirements Document (PRD) or Bug Brief directly in Linear.
2. Plan	/tasks	Analyzes the PRD and generates a detailed Implementation Plan with parent tasks and atomic sub-tasks.
3. Build	/implement	The Builder Agent. Autonomously writes code, runs tests, and commits changes for every task in the plan (using Git Worktrees).
4. Walkthrough	/walkthrough	(Experimental) The Proof Agent. Generates a narrative summary and visual storyboard (screenshots/GIFs) of the changes.
5. Ship	/finalize	Polishes the worktree, resolves conflicts, and opens/updates the GitHub Pull Request.