conductor-pattern_skill

Overview

This skill implements the Conductor pattern to run multiple AI agents (Claude, Codex, Gemini) in parallel using git worktrees. It creates isolated worktrees per agent so each implements the same spec without merge conflicts, then guides PR creation and comparison to assemble a best-of-both merge. It optionally integrates a plan-review step (planview) to approve or iterate on agent plans before execution.

How this skill works

The orchestrator script creates per-agent git worktrees branched from a common base, copies shared config files, and launches agents concurrently (typically via tmux). Each agent works independently in its sandbox, commits its changes to its branch, and the tool helps automate PR creation and comparison. A planview integration can require plan approval before the conductor runs and can annotate plans to modify agent roles or steps.

When to use it

• When you want multiple agents to implement the same specification concurrently and compare outcomes.
• During high-risk or large refactors to hedge risk across agents and increase success probability.
• When separating responsibilities across agents (e.g., one for domain logic, one for tests/types, one for docs).
• When you want plan review and approval before executing parallel agent runs.
• When you need side-by-side PRs to cherry-pick the best pieces and compose a final merge.

Best practices

• Use a planview review loop to approve or refine agent plans before running the conductor.
• Keep shared configuration files (.env, build scripts) synchronized and copied into each worktree.
• Name branches and worktrees consistently (feat/<name>-<agent>) for easy PR creation and comparison.
• Limit each agent’s scope (UI, tests, docs) to avoid redundant conflicting changes.
• Clean up worktrees and temporary branches after merging to avoid git clutter.

Example use cases

• UI redesign: Claude implements structure while Codex refines styling and accessibility.
• API development: Claude writes business logic, Codex produces types and tests, Gemini drafts docs.
• Large refactor: Run Claude and Codex in parallel to hedge implementation risk and pick the best result.
• Automated workflow: Approve plan with planview, then run conductor-planview.sh to execute approved plans.