Parallel AI Coding with Git Worktrees and Custom Claude Code Commands
AI coding is evolving fast. With Claude Code support for custom commands, it's time to upgrade your workflows. One of the most powerful advanced agentic coding techniques is parallel development with Git worktrees—running multiple Claude agents simultaneously on different branches of your codebase using custom slash commands. Our adoption of this technique is inspired by the benchy repository from this video.
Let's break it down step-by-step so you can replicate this advanced workflow in your own repo using custom Claude Code commands.
What is Parallel AI Coding?
Parallel AI coding is an advanced development technique where you run multiple AI agents simultaneously on isolated copies of your codebase to implement the same feature. Using git worktrees and custom Claude Code commands, each agent works independently on its own branch, producing different implementations of the same specification.
This approach leverages the non-deterministic nature of Large Language Models (LLMs) as a feature rather than a limitation. Since the same prompt can produce different valid solutions, parallel AI coding lets you explore multiple solution paths simultaneously and choose the best result.
The key components are:
- Git worktrees for isolated development environments
- Custom Claude Code commands for orchestrating multiple agents
- Parallel execution using Claude's Task tool
- Comparative analysis to select the optimal implementation
Also known as AI imagines, you pick your vision...
Why Parallel Workflows?
Large Language Models (LLMs) are non-deterministic. The same prompt run twice can produce different results. With Claude 4, these results are often good—but they'll be different. Running N parallel agents gives you:
- Redundancy if one agent fails
- Multiple design perspectives
- Better final code by picking the best result
Project Structure
We'll use custom Claude Code commands with git worktrees. Git worktrees allow you to check out multiple branches from the same repository into separate directories - perfect for parallel development. Here's the essential directory structure:
project/
├── .claude/
│ └── commands/
│ ├── init-parallel.md
│ └── exe-parallel.md
├── client/
├── docs/
├── server/
├── specs/
│ └── interview-dashboard.md
└── trees/
├── interview-dashboard-1/
├── interview-dashboard-2/
└── interview-dashboard-3/
Step 1: Set Up Custom Claude Code Commands
Understanding Custom Commands
Claude Code supports custom slash commands that you can create to quickly execute specific prompts or tasks. When you create a command file in .claude/commands/, it becomes available as a slash command with the /project: prefix.
Here's how the system works:
- Command names are derived from the filename (e.g.,
init-parallel.mdbecomes/project:init-parallel) - You can use
$ARGUMENTSplaceholders that get replaced with user input - Commands are version controlled and shareable with your team
- The markdown content becomes the prompt sent to Claude when invoked
For the complete guide on creating custom slash commands, see the official Claude Code documentation.
Let's set this up:
Create the Commands Directory
mkdir -p .claude/commands
Create the Initialization Command
Create .claude/commands/init-parallel.md:
# Initialize Parallel Worktrees
## Variables
FEATURE_NAME: $ARGUMENTS
NUMBER_OF_TREES: $ARGUMENTS
## Instructions
Create NUMBER_OF_TREES git worktrees for parallel development of FEATURE_NAME.
1. Create the trees directory if it doesn't exist
2. For each tree (1 to NUMBER_OF_TREES):
- Create a new git worktree at `trees/FEATURE_NAME-{i}/`
- Create a new branch named `FEATURE_NAME-{i}`
- Copy environment files to each worktree
- Set up development environment in each worktree
Each worktree will be an isolated copy of the codebase on its own branch, ready for independent development.
RUN `mkdir -p trees`
For each worktree:
```bash
git worktree add trees/FEATURE_NAME-1 -b FEATURE_NAME-1
git worktree add trees/FEATURE_NAME-2 -b FEATURE_NAME-2
git worktree add trees/FEATURE_NAME-3 -b FEATURE_NAME-3
```
Copy environment variables and setup each environment:
```bash
cp .env trees/FEATURE_NAME-1/.env 2>/dev/null || true
cp .env trees/FEATURE_NAME-2/.env 2>/dev/null || true
cp .env trees/FEATURE_NAME-3/.env 2>/dev/null || true
```
List the created worktrees:
RUN `git worktree list`
Create the Execution Command
Create .claude/commands/exe-parallel.md:
# Parallel Task Execution
## Variables
PLAN_TO_EXECUTE: $ARGUMENTS
NUMBER_OF_PARALLEL_WORKTREES: $ARGUMENTS
## Run these commands first
RUN `ls -la trees/`
RUN `git worktree list`
READ: PLAN_TO_EXECUTE
## Instructions
We're going to create NUMBER_OF_PARALLEL_WORKTREES new subagents that use the Task tool to create N versions of the same feature in parallel.
This enables us to concurrently build the same feature in parallel so we can test and validate each subagent's changes in isolation then pick the best changes.
The first agent will run in trees/<feature_name>-1/
The second agent will run in trees/<feature_name>-2/
...
The last agent will run in trees/<feature_name>-<NUMBER_OF_PARALLEL_WORKTREES>/
The code in each worktree will be identical to the code in the current branch. It will be setup and ready for you to build the feature end to end.
Each agent will independently implement the engineering plan detailed in PLAN_TO_EXECUTE in their respective workspace.
When each subagent completes their work, have them report their final changes in a `RESULTS.md` file at the root of their respective workspace.
Make sure agents don't run start.sh or any other scripts that would start servers - focus on the code changes only.
Step 2: Create Your Development Plan
Create a clear specification in specs/interview-dashboard.md:
# AI Interview Dashboard Design
## Goal
Design an AI interview interface that handles audio recording, real-time transcription, sentiment analysis, insights extraction, and interview stage management.
## Requirements
- **Audio Recording**: High-quality recording with visual feedback and controls
- **Live Transcription**: Real-time speech-to-text with speaker identification
- **Sentiment Analysis**: Live sentiment tracking with visual indicators
- **Insights Panel**: AI-generated insights, topics, and key moments extraction
- **Interview Stages**: Clear progression through intro, questions, and wrap-up phases
- **Responsive Design**: Works seamlessly on desktop and tablet devices
## Implementation Notes
- Focus on interviewer workflow and cognitive load reduction
- Prioritize accessibility for screen readers and keyboard navigation
- Implement real-time data updates without UI flicker
- Consider mobile fallback for interview participants
- Integrate with existing Agent Interviews design system
## Success Criteria
- Intuitive interviewer experience with minimal learning curve
- Reliable audio recording with clear visual feedback
- Accurate real-time transcription display
- Actionable insights generated during interview flow
- Smooth transitions between interview stages
Step 3: Execute the Parallel Workflow
Now you can use your custom commands to orchestrate parallel development:
Initialize the Worktrees
claude
> /project:init-parallel interview-dashboard 3
This command will:
- Create 3 git worktrees in the
trees/directory - Set up branches
interview-dashboard-1,interview-dashboard-2,interview-dashboard-3 - Copy environment files to each worktree
- Prepare isolated development environments
Execute Parallel Development
> /project:exe-parallel specs/interview-dashboard.md 3
This command will:
- Read your development plan from
specs/interview-dashboard.md - Create 3 parallel subagents using Claude Code's Task tool
- Each agent works independently in their own worktree
- Generate isolated versions of your feature
Step 4: Compare and Test Results
Create a helper script start-parallel-clients.sh to run all versions:
#!/bin/bash
# Start clients for all worktrees in parallel
echo "Starting clients for all worktrees..."
# Colors for output
GREEN='\033[0;32m'
BLUE='\033[0;34m'
YELLOW='\033[1;33m'
NC='\033[0m' # No Color
# Start client in interview-dashboard-1 (port 5174)
echo -e "${BLUE}Starting client in interview-dashboard-1 on port 5174...${NC}"
cd trees/interview-dashboard-1/client && bun run dev &
PID1=$!
echo -e "${GREEN}✓ interview-dashboard-1 client started (PID: $PID1)${NC}"
# Start client in interview-dashboard-2 (port 5175)
echo -e "${BLUE}Starting client in interview-dashboard-2 on port 5175...${NC}"
cd trees/interview-dashboard-2/client && bun run dev &
PID2=$!
echo -e "${GREEN}✓ interview-dashboard-2 client started (PID: $PID2)${NC}"
# Start client in interview-dashboard-3 (port 5176)
echo -e "${BLUE}Starting client in interview-dashboard-3 on port 5176...${NC}"
cd trees/interview-dashboard-3/client && bun run dev &
PID3=$!
echo -e "${GREEN}✓ interview-dashboard-3 client started (PID: $PID3)${NC}"
echo -e "\n${YELLOW}All clients started!${NC}"
echo -e "${YELLOW}Access them at:${NC}"
echo -e " ${GREEN}http://localhost:5174${NC} - interview-dashboard-1"
echo -e " ${GREEN}http://localhost:5175${NC} - interview-dashboard-2"
echo -e " ${GREEN}http://localhost:5176${NC} - interview-dashboard-3"
echo -e "\n${YELLOW}Press Ctrl+C to stop all clients${NC}"
# Wait for any of the processes to exit
wait $PID1 $PID2 $PID3
Run the script:
chmod +x start-parallel-clients.sh
./start-parallel-clients.sh
Step 5: Merge the Best Version
After comparing all versions:
# Navigate to your preferred version
cd trees/interview-dashboard-2
# Review the changes
git status
git diff HEAD~1
# Commit and push
git add .
git commit -m "Interview dashboard design - parallel agent version 2"
git push origin interview-dashboard-2
# Return to main and merge
cd ../..
git checkout main
git merge interview-dashboard-2
Clean Up Worktrees
# List all worktrees
git worktree list
# Remove unused worktrees
git worktree remove trees/interview-dashboard-1
git worktree remove trees/interview-dashboard-3
# Clean up any stale files
git worktree prune
# Remove the trees directory
rm -rf trees/
For more details on git worktree management, see the official git worktree documentation.
Advanced: Managing Custom Commands
Command Organization
You can organize commands in subdirectories:
.claude/commands/
├── parallel/
│ ├── init.md # /project:parallel:init
│ └── execute.md # /project:parallel:execute
├── testing/
│ └── run-suite.md # /project:testing:run-suite
└── deploy.md # /project:deploy
Using Arguments Effectively
Create flexible commands with $ARGUMENTS:
.claude/commands/setup-feature.md:
# Feature Setup
## Variables
FEATURE_NAME: $ARGUMENTS
BRANCH_NAME: $ARGUMENTS
Set up a new feature branch for FEATURE_NAME called BRANCH_NAME with:
- New git branch
- Updated package.json
- Initial component structure
- Test files
Usage:
> /project:setup-feature interview-recording feature/audio-capture
Summary: Advanced Agentic Workflows
This approach gives you:
- Custom slash commands for repeatable workflows
- Variable system for flexible automation
- Parallel subagent execution using Claude Code's Task tool
- Isolated environments with git worktrees
- Multiple implementation perspectives to choose from
This is agentic engineering at scale—not just generating code, but orchestrating multiple AI agents to explore different solutions in parallel.
Huge credit to @disler for his video here video demonstration showing a complete parallel development process.
Final Thoughts
With Claude Code's custom command system and git worktrees, you're not just prompting. You're engineering with AI tools at scale. This workflow gives you compute-level parallelism to ship faster and better.
"Don't just generate code. Multiply AI creativity—and choose the one that matches your vision."
The custom command system makes this workflow repeatable, shareable, and scalable across your entire team.
Happy shipping.
Want to learn more about advanced AI coding workflows? This technique combines Claude Code's custom commands with git's powerful worktree feature to enable true parallel development.