grandamenium

Branch And Prune — Git skill for Claude Code

Git community

Branch and prune: generate N parallel solution variations, review each, prune to a winner, merge with one OK.

How to install Branch And Prune

This entry records only its repository, not the path inside it, so there is no exact command to give. Open grandamenium/branch-and-prune and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Branch And Prune does

Branch and prune: generate N parallel solution variations, review each, prune to a winner, merge with one OK. A self-contained Claude Code skill.

Alternatives in Git

Triage Open Pull Requests — Review, label, and act on all open PRs for a repository using parallel review agents 10.8k ★
Merge Dependabot PRs — @description Evaluate and merge dependabot PRs with parallel builds, dependency-aware batching, and transitive 1.6k ★
GitHub Review PR — Use when a PR needs full review — resolves merge conflicts with the base first, then fixes CI failures, then a 1.5k ★

README

branch-prune

Stop letting your coding agent commit to the first idea it has. Branch and prune: generate N genuinely different solutions in parallel subagents, review each, prune to a winner, merge with one OK.

A Claude Code skill. Self-contained — no external dependencies, no MCP servers, nothing beyond vanilla Claude Code (subagents + git).

Install

mkdir -p ~/.claude/skills/branch-prune
cp SKILL.md ~/.claude/skills/branch-prune/SKILL.md

Per-project install: use `/.claude/skills/branch-prune/SKILL.md` instead.

Use

In any Claude Code session:

/branch

or just ask: "run branch and prune on ". The skill never triggers on its own — you invoke it.

It works at any level: planning a large architecture, building a feature, designing UI, even non-coding tasks like copywriting.

What it does

Asks how many variations you want (no cap; warns about cost on large N).
Suggests 5-10 diversity axes derived from your task — you pick.
Asks how to configure the subagents (forked vs fresh context, model).
Runs one subagent per variation, each in its own git worktree, each with a distinct prompt so outputs genuinely differ. Errors get redeployed, never silently dropped. (After 3 consecutive failures of one variation it asks you what to do — a safety default added by the authors on top of the original "redeploy, no cap" spec, since confirmed by the spec owner as intended behavior.)
Reviews every variation with its own review subagent.
Prunes: a single fair LLM judge with an editable rubric, or you judge manually — your choice.
Delivers per-variation git branches + a report with a winner-based merge suggestion. You give one OK and it merges.

Pro tip

Generate some variations with different LLMs entirely (Gemini, Codex, Claude) and feed them into the judging phase manually. The skill leaves vendor orchestration to you on purpose.

Part of the cortextOS approach to agent engine