Research-Pi: Skill for Claude Code

Research Project Management

Keep the current implementation clear, the project readable and compact, and history in Git. Organize enough to support research; do not turn these principles into a checklist, approval ceremony or recurring report.

Ownership and workspaces

The Leader owns branch and worktree creation, assignment, integration and removal. Executors use their assigned branch and workspace, may commit their own changes within scope, and discuss isolation needs through the available Leader/mailbox channel. They do not independently create, switch, merge or delete branches or worktrees.

Use a worktree only when conflicting parallel implementations, an independent research direction or preservation of a running baseline needs isolation. Multiple executors alone do not require multiple worktrees. Reuse suitable existing workspaces. Place new worktrees under the main project's ignored .worktrees/, never under a user-home scratch directory or nested inside another worktree.

When a temporary workspace's purpose ends, the Leader integrates useful changes or retains a committed branch, then promptly removes the worktree and unneeded branch. First account for active jobs and unique uncommitted or ignored results. Move worthwhile results to their intended project location; do not copy the entire worktree into a backup. Keep a branch only while its research purpose remains useful. Use existing task context and Git to track this, not a new registry.

Git is the version store

Track core code in Git. Before replacing an uncommitted implementation worth retaining, commit the relevant work without sweeping in others' edits. After replacement, delete obsolete code and record the deletion. Recover old implementations from commits or retained branches, not _old, _backup, timestamped source copies, duplicated source directories or commented-out implementations.

Active baselines and algorithm alternatives may coexist when they answer a current research question. Give them meaningful names and reuse shared logic. Once retired, their code belongs in history. Git does not preserve untracked experiment outputs; handle those according to their research value.