Work Forest

A pattern for parallel AI-assisted development

Not a tool—just how I'm working right now. Figured I'd write it down.

The Structure

I have a ~/projects/ folder. Inside it, work-in-progress lives alongside base repos:

~/projects/
├── site-a/                 # base repo (main branch)
├── site-a-add-auth/        # feature branch as folder
├── site-a-fix-nav/         # another feature
├── lib-b/                  # base repo
├── lib-b-extract-theme/    # extracting shared code
└── plans/                  # planning docs (private or public)

Each folder can be a git worktree, a clone, or just a regular checkout. The pattern doesn't require specific tooling.

Planning Documents

Plans live in files with short unique IDs. Context survives session resets:

plans/
├── 3el7f-auth-refactor.md
├── 9xk2m-api-redesign.md
└── w7bn4-mobile-support.md

The ID prefix makes plans greppable across your entire workspace. New sessions read the plan and continue where you left off. Plans can be private, public, or mixed.

Work Clusters

When a task gets big, split it into semantic pieces:

"Build auth. Backend, frontend, database. Work in parallel, then integrate."

AI agents can spawn sub-agents for each piece, working in separate folders, then merge everything.

Stacked PRs

Clean commits, easier reviews:

main
└── refactor-auth     # PR 1: cleanup only
    └── add-oauth     # PR 2: feature on clean code

Helper Scripts

Simple scripts you can vibe-code yourself. You can also create Claude Code skills (.claude/skills/) or add slash commands to CLAUDE.md for workflows you repeat often. See examples for creating plans, managing worktrees, etc.

Trade-offs

This isn't magic. Parallel work means integration overhead. Some work may be redundant. You decide how to split tasks. But it helps me think at a higher level and iterate faster.

View on GitHub